A Multistage Deep Learning Algorithm for

Detecting Arrhythmia
Gokhan ALTAN Novruz ALLAHVERDI Yakup KUTLU
Dept. of Computer Engineering Dept. of Computer Engineering Dept. of Computer Engineering
Iskenderun Technical University KTO Karatay University Iskenderun Technical University
Hatay, Turkey Konya, Turkey Hatay, Turkey
gokhan_altan@hotmail.com novruz.allahverdi@karatay.edu.tr yakupkutlu@gmail.com

Abstract— Deep Belief Networks (DBN) is a deep learning [17] and S transform [7]. Furthermore, PCA [16], [19], [22],
algorithm that has both greedy layer-wise unsupervised and [27], [34], Linear Discriminant Analysis (LDA) [19], [22],
supervised training. Arrhythmia is a cardiac irregularity [34], Independent Component Analysis (ICA) [19], [27] and
caused by a problem of the heart. In this study, a multi-stage qualitative feature selection [4] are the methods that were
DBN classification is proposed for achieving the efficiency of used in feature dimensionality reduction. Finally, selected
the DBN on arrhythmia disorders. Heartbeats from the MIT- features are used to learn the discrimination of the
BIH Arrhythmia database are classified into five groups which arrhythmia heartbeats as a classifier such as artificial neural
are recommended by AAMI. The Wavelet packet networks (ANN) [9], [13], [18], [19], [21], [28]–[30], [33],
decomposition, higher order statistics, morphology and
[34], support vector machines (SVM) [1], [5], [7], [16]–[19],
Discrete Fourier transform techniques were utilized to extract
[25], [27], [33] , binary classifiers [1], the k-Nearest
features. The classification performances of the DBN are
94.15%, 92.64%, and 93.38%, for accuracy, sensitivity, and
Neighbor (k-NN) [3], [6], Linear model-based classifiers
selectivity, respectively. [20], [24], [26], the Gaussian mixture model [5], the Extreme
Learning Machine [22], Hybrid Classifiers [8], [11], [14],
Keywords— Deep learning, Deep Belief Network, [15], [28], the Self-Organization Map [23], Conditional
Arrhythmia, ECG Random Fields [25], Deep Belief Networks (DBN) [8],
[12],and Cluster Analysis [4].
I. INTRODUCTION Deep Learning (DL) is a new and powerful machine
An electrocardiogram (ECG) is one of the biomedical learning algorithm which has begun to be used widely in
signals, which shows the electrical activity of the heart. ECG recent years. DL is used in the assessment of speech
reflects the distribution and reaction of heartbeats and recognition, computer vision, natural language processing
whether the rhythm or the rate of the heart is steady or and biomedical signal applications. DL is a neural network
irregular [1]. ECG is one of the most important methods in model that aims to discover multiple and deeper levels of
the assessment of heart functionality and cardiac diseases processes using multilayer hidden units for a better
such as arrhythmias, congestive heart diseases and more [2]. classification performance. The distinctive characteristics of
DL from NN are having at least two hidden layers [35] and
Arrhythmia is a cardiac irregularity of the heart caused by having a small number of neurons. Fewer neurons often
the problem of tachycardia, bradycardia or non-steady provide convenience to the system by calculating the weights
heartbeats [2]–[4]. Detecting different types of arrhythmia is in the supervised learning stage of the algorithm [36]. These
a very important and vital subject [1], [3]. Many researchers techniques have the ability to train deeper systems with many
have proposed various approaches to the classification of the hidden layers.
arrhythmia beats. The MIT-BIH arrhythmia database (ADB)
is the most widely used in ECG literature. Some techniques In this study, the heartbeats from the ADB are classified
are based on the discrimination between two types of into 5 arrhythmia classes recommended by AAMI standards.
arrhythmias such as premature ventricular contraction (PVC) The moving window analysis technique is used to segment
and non-PVC [5]–[8] and various types of arrhythmias [3], the ECG signals (ECGs) into short-term ECGs that have a
[9]–[17]. Most of the techniques have classified heartbeats length of 341 data points (the R peak is in the center of the
into 5 classes that are defined by the Association for the data). The wavelet packet decomposition (WPD), waveform
Advancement of Medical Instruments (AAMI) [1], [18]–[27] morphology, higher order statics and the discrete Fourier
and with different types of arrhythmias [4], [18], [28]–[30]. transform are used to extract features from segmented short-
In these studies, a variety of features was extracted from term ECGs. Features are classified using the multistage DBN
ECGs and different type classifiers were utilized. The classifier. The aims of this study are to determine the
features have been based on the morphology of efficacy of the DBN classifier and to propose an alternative
waveforms[1], [4], [6]–[8], [13], [15], [20]–[25], [31], DBN based classification model on arrhythmia classification.
temporal information [4], [8]–[12], [15], [17], [20], [21],
[23]–[25], [27], [28], autoregressive modeling coefficients II. MATERIALS AND METHODS
[32], Hidden Markov Modeling [5], Discrete Cosine
Transform (DCT) coefficients [29], [33], high order spectral A. Database
analysis [19], high order statistics (HOS) [3], [6], [7], [25], In the literature, the most used database is the ADB [37]
the Wavelet Transform [15], [21]–[24], [27], [29], [30], the for arrhythmia classification. Therefore, ADB is used in this
Discrete Wavelet Transform (DWT) [7], [14], [19], [22], study. Data supplied to the ADB by the Beth Israel Hospital
[29], QRS geometrical information [14], Principle Arrhythmia Laboratory, contains 48 number of 2-lead ECGs
Component Analysis (PCA) [16], integrate and fire sampler from 25 men aged 32–89, and 22 women aged 23–89; each
modeling [26], Deep belief network on waveforms features has 30min long with 360 Hz sampling frequency.

978-1-5386-4427-0/18/$31.00 ©2018 IEEE

 The AAMI increased the understanding, safety, and The DBN is constructed from two stage learning
effectiveness of medical techniques. Because of AAMI algorithms. The first stage is known as pre-training of the
standardizations on arrhythmia heartbeats such as obtaining network by greedy layer-wise unsupervised training. The
different subjects for training and testing, many supervised number of the RBMs depends on the number of the hidden
classifiers do not perform high classification accuracy rates. layers in the DBN [35]. Each RBM has a visible and a
AAMI standards classify ECG beats into five classes: hidden unit and has a connection with the adjacent RBM.
Normal (N), Ventricular (V), Supraventricular (S), Fusion The visible unit of the first RBM is the input vector of the
(F) and Unknown (Q); beat distributions are explained in [3], model which maps this vector to the representation of the
[25] particularly. hidden unit. The visible units (input vector) of the next
RBMs are the represented hidden units of the previous
B. Preprocessing adjacent RBMs [35]. The details on the DBN functions are
In ECG it is easy to have useless noise because of by given in [36], particularly.
skin-electrode changes or the other recording conditions. RBMs have no visible-visible and hidden-hidden
Noise on the ECGs causes the baseline wander effect. Two connections within units. In this way the parameters of the
median filters are applied to ECGs to handle the baseline DBN such as weights, biases, are evaluated in the
wander effect. Filtered ECGs are segmented into heartbeats unsupervised stage by the aid of the probability of greedy
by using a window which is R peak centered and almost 1 layer-wise method. In the second stage, the obtained
sec in length [3]. classification parameters of the DBN by unsupervised
training are enhanced by the supervised fine-tuning within
C. Feature Extraction the Contrastive Divergence method [35], [40].
In this study, 6 morphological features, 8 from higher The performance of the arrhythmia classification model
order statistics (HOS) of ECGs, 90 from HOS of WPD and is measured by using statistical valuation functions:
46 features from the Fourier transform are extracted. Feature Specificity, Sensitivity, and Accuracy which are obtained
extraction is described in [3], particularly. from the confusion matrix of multistage classification [35].
1) High Order Statistics: The statistical features such as The details on the calculation of performance measures are
given in [3], [35], [36]. TP (True Positive) represents the
mean, median, minimum, and variance usually have an
truly classified heartbeats in the stages of proposed method.
ability to characterize the signals and are frequently used in
machine learning algorithms. But, some signals cannot be
IV. EXPERIMENTAL RESULTS
represented properly by first and second order statistical
features. So, the main stream statistics such as 2nd, 3rd, and The detection of the arrhythmia type is fundamental for a
4th order moments and cumulants are calculated as features. quick and successful treatment. Nowadays, computer-based
2) Wavelet Packet Decomposition: The DWT arrhythmia classification methods have a high accuracy rate
in hard-to-detect arrhythmias and symptoms, as they are
decomposes the coefficients by analyzing low frequency
successful and steady in the diagnosis systems. In this study,
sub-band (approximation) coefficients. However, the WPD a deep learning-based multi-stage classification model is
decomposes both the low frequency sub-band proposed for achieving the efficiency of the DBN for
(approximation) and the high frequency sub-band (detail) automatic arrhythmia classification.
coefficients [3], [38]. Three features are extracted for each
sub-band using HOS (2nd, 3rd, and 4th order) [3]. The dataset, which was used in the study, includes a total
of 150 features: 8 from HOS of ECGs, 90 from 2nd, 3rd, and
3) Morphological Features: QR and RS and slopes of
4th order moments and cumulants of WPT, 6 morphological
right and left sides of R waves are extracted. features, and 46 DFT coefficients. The feature vector is
4) Discrete Fourier Transform (DFT): The DFT is a normalized to 0-1.
practical function that converts a finite sampled time domain
into the list of coefficients of frequency domain. 46 energy A multistage DBN classifier is proposed in this study. N,
S, V, F, and Q types of arrhythmias are classified,
values in the frequency band of 0–50 Hz are calculated [3].
respectively. 4 DBN structures are used in the proposed
III. CLASSIFIER system. Fig. 1 depicts the structure of the proposed classifier.

A. Deep Belief Networks

The aim of the DL is modeling complex, hierarchical and
detailed features in data. DL algorithms are based on
stochastic gradient descent, backpropagation and also new
ideas such as the stacked denoising auto encoder, fine tuning
and more. The DBN is one of the most used probabilistic
generative DL algorithms which consists of stacked
Restricted Boltzmann Machines (RBMs) [35], [39]. The
DBN has multiple layers of hidden units. The top two layers
have undirected connections between them. The lower layers
have the top-down between adjacent units. The DBN has an
ability to make deep assessments of the feature connections
[36].
Fig. 1. Structure of the proposed multistage classification system
 The DBN-based multi-stage automatic arrhythmia Leutheuser et al. compared the real-time classification
classification model consists of 4 stages. Greedy layer-wise systems for arrhythmia detection on mobile devices using
pre-training is used in this model at the unsupervised statistical features, heartbeat features and temporal features
learning stage of the DBN with 10 epochs. The activation to classify 2 types of arrhythmias [6]. They reported their
function of the hidden layers on the supervised learning best accuracy rate of 93.30% using k-NN on android-based
phase is the hyperbolic tangent function to avoid bias in the mobile devices. Alajlan et al. classified two arrhythmia types
gradients and to have a stronger gradient. To unfold the DBN with the SVM system using morphological features, DWT,
to a neural network for the supervised learning stage of the higher order statistics and S transform features. Sensitivity
DBN, model parameters were selected by iterations. We and accuracy are reported as 93.14%, and 93.49%,
experimented only with a limited number of parameters and respectively [7]. Yeh et al. have reported an accuracy rate of
the parameters in which the best classification performance 94.30% using a multistage cluster analysis model with
achieved are given. In the first stage, DBN1 has separated N morphological and shape features to classify five types of
class heartbeats from V, S, F and Q arrhythmia heartbeats. arrhythmias [4]. Yan et al. have reported an accuracy rate of
The DBN1 has 2 hidden layers with 200-530 hidden units. 98.82% with morphological features, heartbeat features and
The output layer has two outputs N and Others (S+V+F+Q). raw two-lead ECGs to classify twelve types of arrhythmias
The learning rate is 4 and the softmax output function was using the DBN [12]. Rahhal et al. fed raw ECG waveforms,
utilized. In the second stage, DBN2 separated S class temporal features and the weights and biases that are trained
heartbeats from V, F and Q arrhythmia heartbeats. The by the DBN as features to the SVM training and
DBN2 has 3 hidden layers with 260-420-120 hidden units. classification [8]. An accuracy rate of 98.49% is reported for
The output layer has two outputs: S and Others (V+F+Q). two types of arrhythmia heartbeats. It is hard to compare the
The learning rate is 4 and the softmax output function was classification performances of the studies, because of the
utilized. In the third stage, DBN3 separated V class different numbers and different types of classified arrhythmia
heartbeats from F and Q arrhythmia heartbeats. The DBN3 heartbeats belonging to different patients. The purposed
has 2 hidden layers with 50-60 hidden units. The output layer multistage DBN classification is based on the arrhythmia
has two outputs V and Others (F+Q). The learning rate is 2 types which are defined by the AAMI. The comparison of
and the sigmoid output function was utilized. In the last the studies using the AAMI is seen in Table 2.
stage, DBN4 separated F class heartbeats from Q arrhythmia
heartbeats. The DBN4 has 3 hidden layers with 150-440-100
hidden units. The output layer has two outputs: V and F TABLE II. COMPARISON OF THE RELATED WORKS ON 5 TYPES OF
ARRHYTHMIAS (N, S, V, F, Q) ACCEPTED BY ANSI/AAMI
arrhythmia heartbeats. The learning rate is 3 and the softmax
output function was utilized. Related Features Classifier Accuracy
Works
The training set of the automatic arrhythmia Martis et al. DCT, PCA ANN 99.12%
classification model includes 3,345 heartbeats from various [33] LS-SVM 89.30%-
PNN 98.17%
types of heartbeat classes and the trained DBN model is 99.52%
tested using 2,542 heartbeat classes which are defined by Martis et al. DWT, ICA, LDA, PCA LS-SVM 87.52%-
AAMI standards. The confusion matrix of the classifier is [19] ANN 97.40%
PNN 98.78%
seen in Table 1. 99.28%
Kim et al. CWT, Morphological ELM 97.94%
[22] feature, DWT, PCA, LDA
TABLE I. CONFUSION MATRIX OF MULTISTAGE CLASSIFIER Chazal et al. Morphological features LD 85.83%
Predicted heartbeats [20]
Labels Tadejko et Morphological features, SOM/LVQ 92.95%
N S V F Q al. [23] Wavelet Transform SVM 97.82%
N 477 2 8 16 4 Llamedo et Wavelet Transform, LD 78.00%
al. [24] Morphological features
True heartbeats

S 0 288 10 3 6 Lannoy et al. Morphological features, W-CRF 85.39%

[25] HBF coefficients,
V 3 0 467 2 3 HOS
Alvarado et Integrate and Fire Sampler LD 93.60%
F 12 8 12 173 9
al. [26] Model, Pulse based
Q 8 2 3 6 478 features
Ye et al. [27] Interval Features, Wavelet SVM 86.40%
Transform, ICA, PCA
Proposed Morphological features, DBN 94.15%
Melin et al. used ANN and a learning vector quantization HOS, WPD, DFT
multistage system to classify fifteen types of arrhythmias a.
CWT: Continuous Wavelet Transform, LD: Linear Discriminant, W-CRF: Weighted Conditional
using fiducial points, segmentation of cycles and Random Fields, ELM: Extreme Learning Machines, PNN: Probabilistic Neural Network, LS-SVM:
Least Square-SVM, SOM/LVQ: Self-Organization Map/Learning Vector Quantization
transformation of cardiac cycles with an accuracy rate of
97.64% to 99.16% [9]. Castillo et al. have presented ANN Martis et al. and Kim et al. have reported high accuracy
with gradient descent and a fuzzy k-NN hybrid intelligent rates in the classification of AAMI types of arrhythmia. Each
system to classify five types of arrhythmias using heartbeat study is based on a different number of heartbeats from
segmentation with an accuracy rate of 98% [28]. Chang et al. different subjects. So it is hard to compare the results in an
used the Gaussian mixture model with hidden Markov model objective way. Enhancing the classification accuracy of
features. Sensitivity, specificity, and accuracy are reported as arrhythmia types is not the first object; the aims of this study
85.71%, 79.82%, and 82.50%, respectively, for six classes of are determining the efficiency of the DBN classifier for
arrhythmias [5]. Martis et al. have reported high arrhythmia classification and proposing a multistage
classification performances using higher order spectral classifier system that is an alternative to other machine
analysis and PCA with ANN and least square SVM [18]. learning models. The classification performances of the
proposed system are 94.15%, 92.64%, and 93.38%, for Melgani, and R. R. Yager, “Deep Learning Approach for
overall accuracy, average sensitivity, and average selectivity, Active Classification of Electrocardiogram Signals,” Inf.
respectively. The results are higher than in most of the Sci. (Ny)., vol. 345, pp. 340–354, Feb. 2016.
related works and this case identifies the success of the DL in
the classification of arrhythmia. [9] 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).,
V. CONCLUSION
vol. 279, pp. 483–497, Sep. 2014.
Various feature extraction methods that characterize the [10] J.-S. Wang, W.-C. Chiang, Y.-L. Hsu, and Y.-T. C. Yang,
irregularity on ECGs, several popular classification methods
“ECG arrhythmia classification using a probabilistic neural
have been investigated for various types of automatic
arrhythmia recognition. The DBN is becoming a popular network with a feature reduction method,”
machine learning algorithm for the classification of Neurocomputing, vol. 116, pp. 38–45, 2013.
automatic arrhythmia. Therefore a DBN based multi-stage [11] H. H. Haseena, A. T. Mathew, and J. K. Paul, “Fuzzy
classification system is constructed. clustered probabilistic and multi layered feed forward
Features are based on a combination of a set of derived neural networks for electrocardiogram arrhythmia
morphological, WPD, Fourier and HOS. A 4 stage DBN classification,” J. Med. Syst., vol. 35, no. 2, pp. 179–188,
based system is utilized in classification of heartbeats. This 2011.
paper presented an attempt to develop a computerized [12] Y. Yan, X. Qin, Y. Wu, N. Zhang, J. Fan, and L. Wang, “A
classifier, which can detect arrhythmia types in an efficient restricted Boltzmann machine based two-lead
and reliable way. The results show that the proposed multi- electrocardiography classification,” 2015 IEEE 12th
stage classifier for discriminating a broad range of heartbeats
International Conference on Wearable and Implantable
performs good with average sensitivity, average selectivity,
Body Sensor Networks (BSN). pp. 1–9, 2015.
and overall accuracy rates of 92.64%, 93.38%, and 94.15%,
respectively. [13] M. Moavenian and H. Khorrami, “A qualitative comparison
of Artificial Neural Networks and Support Vector
ACKNOWLEDGMENT Machines in ECG arrhythmias classification,” Expert Syst.
Appl., vol. 37, no. 4, pp. 3088–3093, 2010.
N.Allahverdi thanks the KTO Karatay University for its
support of this work. [14] M. R. Homaeinezhad, S. A. Atyabi, E. Tavakkoli, H. N.
Toosi, A. Ghaffari, and R. Ebrahimpour, “ECG arrhythmia
REFERENCES recognition via a neuro-SVM-KNN hybrid classifier with
virtual QRS image-based geometrical features,” in Expert
[1] Z. Zhang, J. Dong, X. Luo, K. S. Choi, and X. Wu, Systems with Applications, 2012, vol. 39, no. 2, pp. 2047–
“Heartbeat classification using disease-specific feature 2058.
selection,” Comput. Biol. Med., vol. 46, no. 1, pp. 79–89,
[15] M. Javadi, S. A. A. A. Arani, A. Sajedin, and R.
2014.
Ebrahimpour, “Classification of ECG arrhythmia by a
[2] M. Gabriel Khan, Rapid ECG Interpretation(Contemporary modular neural network based on Mixture of Experts and
Cardiology), 3rd editio. Humana Press, 2007. Negatively Correlated Learning,” Biomed. Signal Process.
[3] Y. Kutlu and D. Kuntalp, “A multi-stage automatic Control, vol. 8, no. 3, pp. 289–296, 2013.
arrhythmia recognition and classification system,” Comput. [16] A. Batra and V. Jawa, “Classification of Arrhythmia Using
Biol. Med., vol. 41, no. 1, pp. 37–45, 2011. Conjunction of Machine Learning Algorithms and ECG
[4] Y. C. Yeh, C. W. Chiou, and H. J. Lin, “Analyzing ECG Diagnostic Criteria,” Int. J. Biol. Biomed., vol. 1, pp. 1–7,
for cardiac arrhythmia using cluster analysis,” Expert Syst. 2016.
Appl., vol. 39, no. 1, pp. 1000–1010, 2012. [17] M. Huanhuan and Z. Yue, “Classification of
[5] P.-C. Chang, J.-J. Lin, J.-C. Hsieh, and J. Weng, Electrocardiogram Signals with Deep Belief Networks,”
“Myocardial infarction classification with multi-lead ECG Computational Science and Engineering (CSE), 2014 IEEE
using hidden Markov models and Gaussian mixture 17th International Conference on. pp. 7–12, 2014.
models,” Appl. Soft Comput., vol. 12, no. 10, pp. 3165– [18] R. J. Martis, U. R. Acharya, K. M. Mandana, a. K. Ray,
3175, 2012. and C. Chakraborty, “Cardiac decision making using higher
[6] H. Leutheuser, S. Gradl, P. Kugler, L. Anneken, M. Arnold, order spectra,” Biomed. Signal Process. Control, vol. 8, no.
S. Achenbach, and B. M. Eskofier, “Comparison of real- 2, pp. 193–203, 2013.
time classification systems for arrhythmia detection on [19] R. J. Martis, U. R. Acharya, and L. C. Min, “ECG beat
Android-based mobile devices,” IEEE Eng. Med. Biol. Soc. classification using PCA, LDA, ICA and Discrete Wavelet
Annu. Conf., vol. 2014, pp. 2690–2693, 2014. Transform,” Biomed. Signal Process. Control, vol. 8, no. 5,
[7] N. Alajlan, Y. Bazi, F. Melgani, S. Malek, and M. A. pp. 437–448, 2013.
Bencherif, “Detection of premature ventricular contraction [20] P. De Chazal and R. B. Reilly, “A patient-adapting
arrhythmias in electrocardiogram signals with kernel heartbeat classifier using ECG morphology and heartbeat
methods,” Signal, Image Video Process., vol. 8, no. 5, pp. interval features,” IEEE Trans. Biomed. Eng., vol. 53, no.
931–942, Jul. 2014. 12, pp. 2535–2543, 2006.
[8] M. M. Al Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. [21] T. Ince, S. Kiranyaz, and M. Gabbou, “A generic and
 robust system for automated patient-specific classification arrhythmia classification using dual tree complex wavelet
of ECG signals,” IEEE Trans. Biomed. Eng., vol. 56, no. 5, based features,” AEU - Int. J. Electron. Commun., vol. 69,
pp. 1415–1426, 2009. no. 4, pp. 715–721, 2015.
[22] J. Kim, S. D. Min, and M. Lee, “An arrhythmia [31] P. D. C. P. De Chazal, M. O’Dwyer, and R. B. Reilly,
classification algorithm using a dedicated wavelet adapted “Automatic classification of heartbeats using ECG
to different subjects.,” Biomed. Eng. Online, vol. 10, no. 1, morphology and heartbeat interval features,” IEEE Trans.
p. 56, 2011. Biomed. Eng., vol. 51, no. 7, pp. 1196–1206, 2004.
[23] P. Tadejko and W. Rakowski, “Hybrid wavelet- [32] D. Ge, N. Srinivasan, and S. M. Krishnan, “Cardiac
mathematical morphology feature extraction for heartbeat arrhythmia classification using autoregressive modeling.,”
classification,” in EUROCON 2007 - The International Biomed. Eng. Online, vol. 1, p. 5, 2002.
Conference on Computer as a Tool, 2007, pp. 127–132. [33] R. J. Martis, U. R. Acharya, C. M. Lim, and J. S. Suri,
[24] M. Llamedo and J. P. Martinez, “Heartbeat classification “Characterization of ECG beats from cardiac arrhythmia
using feature selection driven by database generalization using discrete cosine transform in PCA framework,”
criteria,” IEEE Trans. Biomed. Eng., vol. 58, no. 3 PART 1, Knowledge-Based Syst., vol. 45, pp. 76–82, 2013.
pp. 616–625, 2011. [34] D. Wang and Y. Shang, “Modeling Physiological Data with
[25] G. De Lannoy, D. François, J. Delbeke, and M. Verleysen, Deep Belief Networks.,” Int. J. Inf. Educ. Technol., vol. 3,
“Weighted conditional random fields for supervised no. 5, pp. 505–511, 2013.
interpatient heartbeat classification,” IEEE Trans. Biomed. [35] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle,
Eng., vol. 59, no. 1, pp. 241–247, 2012. “Greedy Layer-Wise Training of Deep Networks,” Adv.
[26] A. S. Alvarado, C. Lakshminarayan, and J. C. Príncipe, Neural Inf. Process. Syst., vol. 19, no. 1, p. 153, 2007.
“Time-based compression and classification of heartbeats,” [36] N. Allahverdi, G. Altan, and Y. Kutlu, “Diagnosis of
IEEE Trans. Biomed. Eng., vol. 59, no. 6, pp. 1641–1648, Coronary Artery Disease Using Deep Belief Networks,” 2.
2012. Int. Conf. Eng. Nat. Sci., p. in press, 2016.
[27] C. Ye, B. V. K. Vijaya Kumar, and M. T. Coimbra, [37] A. L. Goldberger and coworkers, “PhysioNet: a research
“Heartbeat classification using morphological and dynamic resource for studies of complex physiologic and biomedical
features of ECG signals,” IEEE Trans. Biomed. Eng., vol. signals.,” Comput. Cardiol., vol. 27, pp. 179–182, 2000.
59, no. 10, pp. 2930–2941, 2012. [38] M. Y. Gokhale, “Time Domain Signal Analysis Using
[28] O. Castillo, P. Melin, E. Ramírez, and J. Soria, “Hybrid Wavelet Packet Decomposition Approach,” Int’l J.
intelligent system for cardiac arrhythmia classification with Commun. Netw. Syst. Sci., vol. 03, no. 03, pp. 321–329,
Fuzzy K-Nearest Neighbors and neural networks combined 2010.
with a fuzzy system,” Expert Syst. Appl., vol. 39, no. 3, pp. [39] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning
2947–2955, 2012. algorithm for deep belief nets.,” Neural Comput., vol. 18,
[29] H. Khorrami and M. Moavenian, “A comparative study of no. 7, pp. 1527–54, 2006.
DWT, CWT and DCT transformations in ECG arrhythmias [40] Y. Bengio and O. Delalleau, “Justifying and generalizing
classification,” Expert Syst. Appl., vol. 37, no. 8, pp. 5751– contrastive divergence,” Neural Comput., vol. 21, no. 6, pp.
5757, 2010. 1601–1621, 2009.
[30] M. Thomas, M. K. Das, and S. Ari, “Automatic ECG