Accepted Manuscript

Automated Detection of Arrhythmias Using Different Intervals of

Tachycardia ECG Segments with Convolutional Neural Network

U. Rajendra Acharya , Hamido Fujita , Oh Shu Lih ,

Yuki Hagiwara , Jen Hong Tan , Muhammad Adam

PII: S0020-0255(17)30653-9
DOI: 10.1016/j.ins.2017.04.012
Reference: INS 12833

To appear in: Information Sciences

Received date: 9 March 2017

Revised date: 5 April 2017
Accepted date: 7 April 2017

Please cite this article as: U. Rajendra Acharya , Hamido Fujita , Oh Shu Lih , Yuki Hagiwara ,
Jen Hong Tan , Muhammad Adam , Automated Detection of Arrhythmias Using Different Intervals
of Tachycardia ECG Segments with Convolutional Neural Network, Information Sciences (2017), doi:
10.1016/j.ins.2017.04.012

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

 Classification of normal and tachycardia arrhythmias ECG segments.

 Two and five seconds ECG segments are considered.
 Convolutional neural network is employed.
 QRS detection is not performed.
 Accuracy of 92.50% and 94.9% obtained for two and five seconds respectively.

T
IP
CR
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Automated Detection of Arrhythmias Using

Different Intervals of Tachycardia ECG Segments
with Convolutional Neural Network
U. Rajendra Acharya a,b,c, Hamido Fujita d, *, Oh Shu Lih a, Yuki Hagiwara a, Jen Hong Tan a,
Muhammad Adam a

T
IP
a Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore
b Department of Biomedical Engineering, School of Science and Technology, SIM University,

CR
Singapore
c Department of Biomedical Engineering, Faculty of Engineering, University of Malaya,
Malaysia
d

0693 Japan
US
Iwate Prefectural University (IPU), Faculty of Software and Information Science, Iwate 020-
AN
Postal Address: Iwate Prefectural University (IPU), Faculty of Software and Information
*

Science, Iwate 020-0693 Japan

Telephone: +81-19-694-2578; Email Address: [email protected]
M

ABSTRACT
ED

Our cardiovascular system weakens and is more prone to arrhythmia as we age. An arrhythmia
is an abnormal heartbeat rhythm which can be life-threatening. Atrial fibrillation (Afib), atrial
PT

flutter (Afl), and ventricular fibrillation (Vfib) are the recurring life-threatening arrhythmias that
affect the elderly population. An electrocardiogram (ECG) is the principal diagnostic tool
employed to record and interpret ECG signals. These signals contain information about the
CE

different types of arrhythmias. However, due to the complexity and non-linearity of ECG
signals, it is difficult to manually analyze these signals. Moreover, the interpretation of ECG
signals is subjective and might vary between the experts. Hence, a computer-aided diagnosis
AC

(CAD) system is proposed. The CAD system will ensure that the assessment of ECG signals is
objective and accurate. In this work, we present a convolutional neural network (CNN)
technique to automatically detect the different ECG segments. Our algorithm consists of an
eleven-layer deep CNN with the output layer of four neurons, each representing the normal
(Nsr), Afib, Afl, and Vfib ECG class. In this work, we have used ECG signals of two seconds and
five seconds’ durations without QRS detection. We achieved an accuracy, sensitivity, and
specificity of 92.50%, 98.09%, and 93.13% respectively for two seconds of ECG segments. We
obtained an accuracy of 94.90%, the sensitivity of 99.13%, and specificity of 81.44% for five
 ACCEPTED MANUSCRIPT

seconds of ECG duration. This proposed algorithm can serve as an adjunct tool to assist
clinicians in confirming their diagnosis.

Keywords – Arrhythmia, atrial fibrillation, atrial flutter, convolution neural network, deep
learning, electrocardiogram signals, ventricular fibrillation.

T
1. Introduction

IP
CR
According to the report by the United Nations in 2015, the world is facing an aging population
[33]. It is estimated that the number of people aged 60 years and above will grow by 56.00%
from 901 million to 1.4 billion by 2030. Furthermore, the growing population (60 years and older)

US
is expected to be double by 2050, rising to nearly 2.1 billion [33]. The increase in elderly
population poses economic [3, 33] and health care issues [28, 33] to the world. Our
AN
cardiovascular system grows weaker and becomes more receptive to diseases as we grow older
[29]. Moreover, the arteries stiffen and muscle wall of the left ventricle thickens with aging,
resulting in a decrease in the compliance of blood vessels of the arteries [6]. Consequently, it
M

affects the overall function of the heart which leads to arrhythmia. Hence, arrhythmia is one of
the health conditions that the elderly subjects encounter in the society [4, 6]. Arrhythmia is
ED

defined as the abnormal rhythm of the heartbeat which can be harmless or critical. The atrial
fibrillation (Afib), atrial flutter (Afl), and ventricular fibrillation (Vfib) are the recurrent types of
arrhythmias reported in the elderly [6].
PT

The Afib is a commonly occurring arrhythmia caused due to various health complications.
During Afib, the contraction of the atria is asynchronous due to the fast firing of electrical
CE

impulses from several parts of cardiac re-entry [2]. Re-entry occurs when an impulse fails to die
out after normal activation of the heart and continues to re-excite the heart. In fact, the
AC

electrocardiogram (ECG) rhythm of Afib is fast and beating at a rate of 150 to 220 beats in a
minute. It has an abnormal R-R interval, irregular and fast ventricular contraction, and P wave
is absent in the ECG signal [29]. In Afl, the atrial contracts rapidly between 240 and 360 beats per
minute and have a replicating saw-tooth waveform, known as flutter wave. Afl occurs when the
atria undergo chaotic electrical signals [2]. Vfib is usually caused by rapid heartbeat known as
ventricular tachycardia (VT). This abnormal heartbeat is due to abnormal electrical impulses in
the ventricles. During this, ventricles contract chaotically and haphazardly. It can be seen in the
 ACCEPTED MANUSCRIPT

ECG morphology, which records an unrefined and erratic fluctuation of ECG signal with the
absence of QRS complex wave [29]. Typical plots of Nsr, Afib, Afl, and Vfib ECG signals are shown
in Figure 1 and Figure 2.

Therefore, the morphology of ECG signals contains vital details about the conditions of the
heart. Thus, the ECG signal is beneficial in the detection and diagnosis of cardiac health [2].
However, ECG signals are highly nonlinear and any changes in the ECG signals is difficult to

T
observe as its amplitude is in millivolts [2, 13]. Further, the indications of cardiac abnormalities

IP
are faithfully indicated in the ECG signals during 24-hour Holter recording. Thus, manual
interpretation of the ECG signals can be time-consuming, taxing and subjective due to the long

CR
recordings. Moreover, there is a great possibility that important information captured in the
ECG morphology may be overlooked. Hence, a computer-aided diagnosis (CAD) system can be
employed to reduce subjective variabilities in the diagnosis and reduce the time taken to
analyze the ECG signals [25].
US
Table 8 shows the studies conducted on CAD system to automatically detect arrhythmias and
AN
categorize different types of arrhythmia into their respective classes. Wang et al. [34] performed
short-time multifractal characterization of Afib, Vfib, and VT classes of ECG beats and recorded
an accuracy of 99.40% for Afib, 97.20% for Vfib and 97.80% for VT using fuzzy Kohonen network
M

classifier. Martis et al. [26, 27] have conducted a three-class study to automatically diagnose Afib,
Afl, and Nsr ECG signals. In [27], they have employed higher order spectra methods on 641 Nsr,
ED

855 Afib, and 887 Afl ECG beats. Then these ECG beats are subjected to independent component
analysis (ICA) to select highly significant features. Their method yielded an accuracy, sensitivity,
PT

and specificity of 97.65%, 98.16%, and 98.75% with the k-nearest neighbor classifier. In their
another study [26], they performed a discrete cosine transform combined with ICA on the ECG
beats. Their proposed technique attained an average accuracy of 99.45%.
CE

In addition, Fahim et al. [10] employed a data mining approach with expectation-maximization-
based clustering on 50 compressed ECG signals obtained from an open-source database. They
AC

used correlation-based feature subset selection technique to reduce the number of features.
Then, the selected features are fed into the classifier. They detected Afib, premature ventricular
contraction, and Vfib with an average accuracy of 97.00% using the rule-based system. Acharya
et al. [1] proposed a CAD system to automatically detect and identify same four ECG classes
(Nsr, Afib, Afl, Vfib) using the entire database (614,526 ECG beats) obtained from an open-source
database [14]. They extracted entropy features from the ECG signals. These extracted features
 ACCEPTED MANUSCRIPT

were subjected to feature reduction and the selected 14 significant features were fed into the
decision tree classifier, yielding an accuracy of 96.30%, sensitivity of 99.30%, and specificity of
84.10%. Further, Desai et al. [8] also implemented a CAD system to diagnose the four-class
arrhythmia (Afib, Afl, Nsr, and Vfib). However, they used a smaller dataset (3,858 ECG beats)
obtained from the same open-source database [14] in their work. They applied the recurrence
quantification analysis parameters to the ECG beats. Then, the features are arranged according
to the F-value index. They achieved an accuracy of 98.37% with the rotation forest classifier.

T
IP
However, from the literature [1, 8, 10, 26, 27], it can be noted that these CAD systems have a
standardized workflow whereby the signals are pre-processed first, then segmented. Then the

CR
signals are subjected to features extraction, followed by feature selection to select only
significant features for classification. In this study, we did not follow the conventional process
of an automated CAD system. This is unlike the previous works recorded in Table 8 as no

US
features extraction or selection is implemented in this work. We employed an eleven-layer
convolutional neural network (CNN) to automatically classify the four classes of ECG signals
AN
(Nsr, Afib, Afl, and Vfib). Hence, in this study, there is no need to experiment with different
features extraction techniques or determine which classifier performs the best with the extracted
features.
M

CNN has recently been employed in the automated classification of ECG signals. Kiranyaz et al.
[21] studied the patient-specific ECG categorization and monitoring system using three-layer
ED

CNN with only R-peak wave. They attained an accuracy of 97.60% and 99.00% in the detection
of supraventricular ectopic beats and ventricular ectopic beats respectively. Zubair et al. [36]
PT

used CNN with 44 recordings of ECG signals obtained from MIT-BIH database. They extracted
R-peak ECG beat patterns for the training of the three-layer CNN. They achieved 92.70%
accuracy in detecting the ECG beats into their respective classes (normal, fusion beat,
CE

supraventricular ectopic beat, unknown beat, and ventricular ectopic beat). These works [21, 36]
detected QRS wave in their automated classification. Nevertheless, in our study, no detection of
AC

QRS wave is implemented.

2. Data Used
 ACCEPTED MANUSCRIPT

In this work, the ECG signals were obtained from a publicly available arrhythmia database. We
have obtained Vfib (Ventricular Fibrillation) ECG signals from Creighton University ventricular
tachyarrhythmia, Afib (Atrial Fibrillation) and Afl (Atrial Flutter) ECG signals from MIT-BIH
atrial fibrillation, and Afib (Atrial Fibrillation), Afl (Atrial Flutter), and Nsr (Normal Sinus Rhythm)
ECG signals from MIT-BIH arrhythmia database [14]. In this work, we have used lead II ECG
signals.

T
The details of the ECG signals used in this study is shown in Table 1. We have used two

IP
different durations of ECG segments (two seconds and five seconds) in this work. The total
number of ECG segments used for net A (two seconds) and net B (five seconds) is 21,709 and

CR
8,683 respectively.

Table 1
Overview of the data used in this study.
Database Data Used
US
AN
Afib,
MIT-BIH Atrial Fibrillation (afdb)
Afl
Afib,
M

MIT-BIH Arrhythmia (mitdb) Afl,

Nsr
Creighton University Ventricular Tachyarrhythmia
ED

Vfib
(cudb)
PT

Table 2
Overview of the ECG segments (two and five seconds) used in this study.
Type Number of Segments Number of Segments
CE

(Two Seconds) (Five Seconds)

(Net A) (Net B)
Nsr 902 361
AC

Afib 18,804 7,521

Afl 1,840 736
Vfib 163 65
Total segments 21,709 8,683
 ACCEPTED MANUSCRIPT

3. Methodology
3.1 Pre-processing

The ECG signals from Creighton University ventricular tachyarrhythmia and MIT-BIH atrial
fibrillation are sampled at a frequency of 250 Hz whereas the ECG signals acquired from MIT-
BIH arrhythmia are sampled at a frequency of 360 Hz. Hence, the ECG signals from MIT-BIH

T
arrhythmia database are downsampled from 360 to 250 Hz. Then, all the ECG signals are
denoised and the baseline is removed with Daubechies wavelet 6 [31].

IP
Further, the ECG signals are segmented and sorted according to the cardiac conditions with the

CR
prescribed annotations retrieved from the public database. In this study, we segmented the ECG
signals of four classes into net A and net B without any wave detection. Each segment is

US
normalized with Z-score normalization to address the problem of amplitude scaling and to
eliminate the offset effect before we feed the ECG segments into the 1-dimensional deep
learning CNN for training and testing. An illustration of two seconds (net A) and five seconds
AN
(net B) ECG segments used in this work are shown in Figure 1 and Figure 2 respectively.
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

Figure 1. An illustration of ECG segments for net A.

PT
CE
AC
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

Figure 2. An illustration of ECG segments for net B.

PT

3.2 Convolutional Neural Network (CNN)

CE

Convolutional neural network (CNN) is first introduced by Fukushima in 1980 [11] and later
AC

improved by LeCun et al [23]. It is a form of deep learning where the structure is made up of
many hidden layers and parameters [23]. Further, the CNN can self-learn and self-organize
which does not require supervision [11]. CNN has been applied in diverse applications such as
object recognition [30], image classification [22], and handwriting classification [7]. It is also
employed in the medical field as an automated diagnostic tool to aid clinicians [16, 17, 18, 32,
35].
 ACCEPTED MANUSCRIPT

It is noted that CNN eliminates the need for pre-processing and separate feature extraction
technique [24]. Therefore, it can help to reduce the burden during training and selecting the best
feature extraction technique for the automated detection of arrhythmias. Also, there is a
possibility of attaining better performance if we can achieve a fitting learning based on the
trained hidden layers by learning the structure of the data. Thus, we used CNN in this study for
these reasons.

T
IP
3.3 The Architecture

CR
The primary operations involved in CNN are convolution, non-linearity, pooling, and
classification [15, 18]. Two architectures of CNN (net A and net B) are proposed in this work.

US
Figure 3 illustrates the working architecture of net A with 500 input samples. The architecture
for net B is illustrated in Figure 4 with 1,250 input samples.
AN
Stride refers to the number of samples the filter matrix slides over the input matrix. Therefore,
in this work, we have used 1 and 2 strides (see Table 3 and Table 4). When the stride is 1, the
filter is moved from one sample to another at a time and when the stride is 2, the filter moves 2
M

samples at a time. A bigger stride will result in smaller feature maps and vice versa.

For both net A and net B, the input layer (layer 0) is convolved with a kernel size of 27 to
ED

produce layer 1. A max-pooling of size 2 is applied onto every feature map (layer 2). Then, the
feature maps from layer 2 are convolved with a kernel size of 14 (net A) and 15 (net B)
PT

respectively to obtain layer 3. A max-pooling of size 2 is again applied to every feature map
(layer 4). The feature maps from layer 4 are then convolved with a kernel size of 3 (net A) and 4
(net B) to produce layer 5 in net A and net B respectively. A max-pooling of size 2 is applied
CE

onto every feature map (layer 6). Then, the feature maps from layer 6 are once again, convolved
with a kernel size of 4 (net A) and 3 (net B) to obtain layer 7 for both net A and net B
AC

accordingly. A max-pooling of size 2 is again applied to every feature map (layer 8). Finally, the
neurons of every feature maps in layer 8 are fully connected to 30 neurons in layer 9, which is
also fully connected to 10 and 4 outputs in layers 10 and 11 respectively.

The leaky rectifier linear unit [19] is used as an activation function for layer 1, 3, 5, 7, 9, and 10.
We have used the softmax function for the last layer (layer 11) and Xavier initialization [12] for
the weights of layers 1, 3, 5, 7, 9, and 10.
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M

Figure 3. The architecture of the proposed CNN for net A.

ED

Table 3
PT

The details of CNN structure for net A.

Number of Neurons Kernel Size for Each
Layers Type Stride
(Output Layer) Output Feature Map
CE

0-1 Convolution 474 x 3 27 1

1-2 Max-pooling 237 x 3 2 2

2-3 Convolution 224 x 10 14 1

AC

3-4 Max-pooling 112 x 10 2 2

4-5 Convolution 110 x 10 3 1

5-6 Max-pooling 55 x 10 2 2

6-7 Convolution 52 x 10 4 1

7-8 Max-pooling 26 x 10 2 2

8-9 Fully-connected 30 - -
 ACCEPTED MANUSCRIPT

9-10 Fully-connected 10 - -

10-11 Fully-connected 4 - -

T
IP
CR
US
AN
Figure 4. The architecture of the proposed CNN for net B.
M

Table 4
The details of CNN structure for net B.
Number of Neurons Kernel Size for Each
Layers Type Stride
ED

(Output Layer) Output Feature Map

0-1 Convolution 1224 x 3 27 1

1-2 Max-pooling 612 x 3 2 2

PT

2-3 Convolution 598 x 10 15 1

3-4 Max-pooling 299 x 10 2 2

CE

4-5 Convolution 296 x 10 4 1

5-6 Max-pooling 148 x 10 2 2

6-7 Convolution 146 x 10 3 1

AC

7-8 Max-pooling 73 x 10 2 2

8-9 Fully-connected 30 - -

9-10 Fully-connected 10 - -

10-11 Fully-connected 4 - -
 ACCEPTED MANUSCRIPT

3.4 Training

Standard backpropagation [5] with a batch size of 10 is implemented for stochastic learning. The

T
weights are updated according to equation (1).

IP
( )

CR
(1)

where w, l, n, 𝝀, ts, x, and c denotes the weight, layer number, learning rate, regularization

US
parameter, total number of training samples, batch size, and cost function respectively. In
addition, the biases are updated through equation (2).
AN

(2)
M

In this work, we have used learning rate, regularization, and momentum parameters. The
parameters are set at 0.002, 0.2, and 0.7 respectively.
ED

3.5 Testing
PT

After every round of training epoch is completed, our algorithm performs a test on the CNN
CE

model. We used 30% of the training set (90%) for validation of the algorithm after every epoch.
A total of twenty epochs of training and testing iterations was run. An illustration of the
AC

distribution of ECG segments used for training and testing procedures can be seen in Figure 5.
 ACCEPTED MANUSCRIPT

T
IP
CR
Figure 5. The distribution of ECG segments used for training and testing.

3.6 k-fold Cross-validation

US
AN
We have used ten-fold cross-validation [9] in this work. Therefore, the total ECG segments for
net A (21,709 segments) and net B (8,683 segments) are divided into ten equal portions. Nine out
of ten portions are used for training and the rest (one-tenth) of the ECG segments are used for
M

testing. This procedure is repeated ten times by shifting the testing data portion. In each fold,
the performances namely the specificity, sensitivity, and accuracy) are evaluated. The average
ED

of all ten-folds gives the total performance of the system.

PT

4. Results
CE

We have trained our algorithm on a workstation with two Intel Xeon 2.40 GHz (E5620)
processor and a 24GB RAM. It took an average of 557.812 seconds to complete an epoch of
AC

training for net A and 256.332 seconds for net B.

Table 5 and Table 6 show the confusion matrix for two-second and five-second segment
respectively. It can be seen from Table 5 that; 93.13% ECG segments are correctly classified as
Nsr class. 92.89% of ECG segments are correctly classified as Afib. A total of 8.64% Afl ECG
segments is wrongly classified as Nsr, Afib, and Vfib. Furthermore, more than a third of Vfib is
wrongly classified as Afib.
 ACCEPTED MANUSCRIPT

Also in Table 6, 18.56% of Nsr ECG segments are wrongly classified as Afib and Afl. Further, 7.11%
Afib segments are incorrectly classified as Nsr, Afl, and Vfib ECG segments. Out of 736 Afl ECG
segments, 86.96% are accurately classified as Afl. Again, more than a third of the Vfib segments
are wrongly classified as Afib.

The overall classification results for net A and net B is tabulated in Table 7. An accuracy of 92.50%
and a sensitivity and specificity of 98.09% and 93.13% respectively is achieved using net A. Also,

T
an average accuracy of 94.90%, and a sensitivity and specificity 99.13%, and 81.44% respectively

IP
are obtained for net B.

Table 5

CR
Confusion matrix for net A.
Original/
Nsr Afib Afl Vfib Acc (%) PPV (%) Sen (%) Spec (%)
Predicted
Nsr
Afib
Afl
840
363
32
45
17,467
115
17
597
1,681
0
377
12
97.88
92.82
96.41
US
67.85
98.75
73.02
93.13
92.89
91.36
98.09
92.39
96.87
AN
Vfib 3 61 7 92 97.88 19.13 56.44 98.19
* Acc = Accuracy, PPV = Positive Predictive Value, Sen = Sensitivity, Spec = Specificity
M

Table 6
Confusion matrix for net B.
Original/
Nsr Afib Afl Vfib Acc (%) PPV (%) Sen (%) Spec (%)
ED

Predicted
Nsr 294 55 12 0 98.40 80.33 81.44 99.13
Afib 57 7,289 116 59 95.32 97.67 96.92 85.03
PT

Afl 15 75 640 6 97.37 82.90 86.96 98.34

Vfib 0 44 4 17 98.70 20.73 26.15 99.25
CE

Table 7
The overall classification results for the classification of Nsr, Afib, Afl, and Vfib classes.
AC

Segment
TP TN FP FN Acc (%) PPV (%) Sen (%) Spec (%)
Length
Two
20,409 840 62 398 92.50 99.70 98.09 93.13
seconds
Five
8,250 294 67 72 94.90 99.19 99.13 81.44
seconds
*TP = True Positive, TN = True Negative, FP = False Positive, FP = False Negative
 ACCEPTED MANUSCRIPT

5. Discussion

The number of Vfib segments (Table 2) used in this work are too few (163 and 65 ECG segments
in net A and net B respectively) and hence resulted in low sensitivity and PPV. Hence, the
performance of CNN gets affected by the number of subjects (data) used in each class.

In this work, net B (five seconds long ECG signal) performed slightly better than net A (two

T
seconds long ECG signal) as there are additional three seconds of additional information on

IP
ECG morphology. However, the results of two (two and five second) time durations are
comparable.

CR
Also, CNN is invariant to translation. Therefore, in this work, the ECG segments are not
affected by time shifting and scaling thus there is no need to perform QRS detection in the pre-

US
processing stage. Normally, the primary steps involved in analyzing ECG signals are (i) filtering
of noise, (ii) detection of QRS complex, (iii) extraction of R-peak, and (iv) formulation of feature
AN
set [2]. Nonetheless, we did not implement step (ii) and (iii) in this work. Most of the works
reported in Table 8 have detected QRS wave in their study. Our results for net A and net B are
comparable to the previous works reported (in Table 8) which proves that the detection of QRS
M

wave is not necessary for the classification of arrhythmia.

In addition, the sensitivity rate achieved for net A (98.09%) and net B (99.13%) is comparable to
ED

those studies summarized in Table 8. Our group [1] obtained a sensitivity of 99.30% using a
total of 614,526 ECG beats (75,815 Nsr beats, 520,292 Afib beats, 14,257 Afl beats, and 4,162 Vfib
beats). In this present work, we obtained a sensitivity of 98.09% and 99.13% for two and five
PT

seconds’ durations with a total of 21,709 and 8,683 ECG segments for net A and net B
respectively.
CE

Additionally, in contrast to the authors [22, 24, 8, 1] in Table 8, we analyzed the ECG signals in
short-term duration (two-second and five-second segments) instead of analyzing one beat of
AC

ECG signal. Normally, doctors analyze a short-duration of ECG signals, not just an ECG beat
for diagnosis. Therefore, it is more realistic to feed two and five seconds of ECG signals to the
CNN structure for the automated detection of arrhythmias. Hence, in this study, we segmented
our ECG signals into two-second and five-second ECG segments.
 ACCEPTED MANUSCRIPT

Table 8
Summary of selected studies conducted for the detection of arrhythmia using the same database.
Author, Year Database Special Characteristics ECG Rhythms Classifier Performance
Three-Class
Afib:
Acc = 99.40%
 No QRS detection Sen = 98.30%

T
performed Spec = 100.00%

IP
 Analysis of 1.2 second
ECG segment Vfib:
Wang et al., Afib,

CR
 Analysis of 1.8 second Fuzzy Kohonen Acc = 97.20%
2001 mitdb Vfib,
ECG segment network Sen = 98.30%
[34] VT
 Analysis of 2.4 second Spec = 96.700%
ECG segment
 Two-layer fuzzy
Kohenen network
US VT:
Acc = 97.80%
AN
Sen = 95.00%
Spec = 99.20%
 QRS detection
Martis et al., Afib, Acc = 99.50%
afdb, performed K-nearest
M

2013 Afl, Sen = 100.00%

mitdb  Analysis of one ECG neighbor
[27] Nsr Spec = 99.22%
beat (2,383 beats)
 QRS detection
ED

Martis et al., Afib, Acc = 99.45%

afdb, performed K-nearest
2014 Afl, Sen = 99.61%
mitdb  Analysis of one ECG neighbor
[26] Nsr Spec = 100.00%
beat (2,942 beats)
PT

Four-Class
Afib,
CE

Atrial

 QRS detection premature

Fahim et al., performed beat,
AC

MIT-BIH
2011  Analysis of ten-second Premature Rule-based Acc = 97.00%
physiobank
[10] ECG segment (800 ventricular (average)
segments) contraction,
Vfib or Vfl

Acharya et afdb,  QRS detection Afib, Acc = 96.30%

Decision tree
al., 2016 cudb, performed Afl, Sen = 99.30%
 ACCEPTED MANUSCRIPT

[1] mitdb  Analysis of one ECG Vfib, Spec = 84.10%

beat (614,526 beats) Nsr

 QRS detection Afib,

Desai et al., afdb,
performed Afl,
2016 cudb, Rotation forest Acc = 98.37%
 Analysis of one ECG Vfib,
[8] mitdb
beat (3,858 beats) Nsr
 No QRS detection

T
performed
 Analysis of two- Net A:

IP
second ECG segment Acc = 92.50%
(21,709 segments) Sen = 98.09%

CR
Afib,
afdb,  Analysis of five- Spec = 93.13%
Current Afl, Convolutional
cudb, second ECG segment
study Vfib, neural network
mitdb (8,683 segments) Net B:
 No feature extraction
or feature selection
involved
US
Nsr
Acc = 94.90%
Sen = 99.13%
Spec = 81.44%
AN
 Eleven-layer deep
CNN
*Acc = Accuracy, Sen = Sensitivity, Spec = Specificity
M

*Afib = Atrial fibrillation, Afl = Atrial flutter, Vfib = Ventricular flutter, Vfl = Ventricular flutter, VT = Ventricular
tachycardia, Nsr = Normal sinus rhythm
*afdb = MIT-BIH atrial fibrillation, cudb = Creighton university ventricular tachyarrhythmia, mitdb = MIT-BIH
arrhythmia
ED

It is evident that our proposed algorithm is more robust as compared to the rest of the works
PT

mentioned in Table 8. Overall, our proposed system does not require any QRS detection. Also
in this work, the feature extraction and selection and classification are merged into one single
model. Furthermore, we have validated the performance of our deep learning model in this
CE

work using net A and net B ECG segments.

To the best of our knowledge, this is the first study to implement an eleven-layer CNN for the
AC

automated detection system of Afib, Afl, Nsr, and Vfib ECG signals without the detection of QRS
complex.

The main highlights of our proposed algorithm are as follows:

i. CNN is invariant to translation, therefore, no pre-processing of QRS detection is needed

in this work.
 ACCEPTED MANUSCRIPT

ii. No QRS detection is required in this work.

iii. Feature extraction, feature selection, and classification steps are merged in the CNN
algorithm.
iv. Ten-fold cross-validation is used for the evaluation of CNN performance in this work.
Hence, the reported performance is robust.

The drawbacks of our proposed algorithm are as follows:

T
i. Requires a lot of data (big data) for training.

IP
ii. Takes more time to train the data.

CR
6. Conclusion

US
Generally, the presence of arrhythmia is reflected in the ECG morphology. Essentially, with
AN
many elderly affected by serious arrhythmias, there is a need to design an efficient and robust
CAD system to accurately and automatically detect various types of arrhythmias. In this work,
we have developed a CNN to automatically classify the four classes (Nsr, Afib, Afl, and Vfib) using
M

21,709 ECG segments of net A and 8,683 ECG segments of net B. Our proposed algorithm
achieved an accuracy of 92.50% and a sensitivity and specificity 98.09%, and 93.13% respectively
ED

for net A. Also, we obtained an average accuracy of 94.90% and a sensitivity and specificity
99.13%, and 81.44% respectively for net B. Hence, it is evident that our developed system has
potential to be implemented in clinical settings. Our proposed toolkit can serve as an adjunct
PT

tool to assist the clinicians to cross-check their findings. Moreover, clinicians can recommend
appropriate treatments promptly and avoid further deterioration of cardiac condition. Further,
CE

the robustness of the proposed system can be improved by using large arrhythmia database
with more number of Vfib, Afib, Afl, and Nsr ECG segments. In future, we intend to use a huge
database and employ the Keras models [20] for the validation of the CNN instead of k-fold
AC

cross-validation strategy. The performance of our method is slightly lower than results of few
other methods reported in Table 8. It is because in our work, we have used blind-fold validation
and works in the table have used ten-fold cross validations. But however, we intend to improve
the performance of our proposed model by using (i) more number of samples in each class, (ii)
data augmentation, and (iii) bagging algorithm. We will be exploring the possibility of using
this system to diagnose other cardiac classes like, myocardial infarction and coronary artery
 ACCEPTED MANUSCRIPT

diseases. We propose to automatically classify the ECG signals using CNN without performing
any noise filtering in our future work.

7. References

T
1. U. R. Acharya, H. Fujita, M. Adam, S. L. Oh, J. H. Tan, V. K. Sudarshan, J. E. W. Koh.

IP
Automated characterization of Arrhythmias using nonlinear features from tachycardia
ECG beats. IEEE International Conference on Systems, Man, and Cybernetics, 2016.

CR
2. U. R. Acharya, J. S. Suri, J. A. E. Spaan, S. M. Krishnan. Advances in Cardiac Signal
Processing. New York, Springer-Verlag Berlin Heidelberg, 2007.
3. S. Bawany. This is the economic impact of an aging Singaporean workforce. Singapore

US
Business Review, 2013. (Last accessed: 24 February 2017).
4. C. Berry, A. C. Rankin, A. J. B. Brady. Bradycardia and tachycardia occurring in older
AN
people: An introduction. British Journal of Cardiology 11(1), 2004.
5. J. Bouvrie. Notes on convolutional neural network, 2007.
6. G. V. Chow, J. E. Marine, J. L. Fleg. Epidemiology of arrhythmias and conduction
M

disorders in older adults. Clinics in Geriatric Medicine 28(4): 539-553, 2012.

7. D. C. Ciresan, U. Meier, L. M. Gambardella, J. Schmidhuber. Convolutional neural
ED

network committees for handwritten character classification. IEEE International

Conference on Document Analysis and Recognition, 1135-1139, 2011.
8. U. Desai, R. J. Martis, U. R. Acharya, C. G. Nayak, G. Seshikala, S. K. Ranjan. Diagnosis
PT

of multiclass tachycardia beats using recurrence quantification analysis and ensemble

classifiers. Journal of Mechanics in Medicine and Biology, 16(1), 2016.
CE

9. R. O. Duda, P. E. Hart, D. G. Stork. Pattern Classification 2nd Edition. New York, John
Wiley and Sons, 2001.
10. S. Fahim, I. Khalil. Diagnosis of cardiovascular abnormalities from compressed ECG: A
AC

datamining-based approach. IEEE Transactions on Information Technology in

Biomedicine, 15: 33-39, 2011.
11. K. Fukushima. Neocognitron: a self-organizing neural network model for a mechanism
of pattern recognition unaffected by shift in position. Biological Cybernetics 36: 193-202,
1980.
 ACCEPTED MANUSCRIPT

12. X. Glorot, Y. Bengio. Understanding the difficulty of training deep feedforward neural
networks. Aistats, 2010.
13. A. L. Goldberger. Clinical Electrocardiography: a simplified approach. Mosby, St. Louis,
MO, USA, 2012.
14. A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. H. Ivanov, R. G. Mark,
J. E. Mietus, G. B. Moody, C. K. Peng, H. E. Stanley. PhysioBank, PhysioToolkit, and
PhyisoNet: Components of a new research resource for complex physiologic signals:

T
Circulation 101(23): e215-e220, 2000.

IP
15. I. Goodfellow, Y. Bengio, A. Courville. Deep learning. MIT Press,

CR
http://www.deeplearningbook.org, 2016.
16. V. Golkov, A. Dosovitskiy, J. I. Sperl, M. I. Menzel, M. Czisch, P. Sämann, T. Brox, D.
Cremers. q-Space deep learning: twelve-fold shorter and model-free diffusion MRI scans.

US
IEEE Transactions on Medical Imaging 35(5): 1344-1351, 2016.
17. M. J. J. P. van Grinsven, B. van Ginneken, C. B. Hoyng, T. Theelen, C. I. Sánchez. Fast
convolutional neural network training using selective data sampling: application to
AN
hemorrhage detection in color fundus images. IEEE Transactions on Medical Imaging
35(5): 1273-1284, 2016.
M

18. N. Hatipoglu, G. Bilgin. Cell segmentation in histopathological images with deep

learning algorithms by utilizing spatial relationships. Medical and Biology in
Engineering and Computing, 1-20, 2017.
ED

19. K. He, X. Zhang, S. Ren, J. Sun. Delving deep into rectifiers: Surpassing human-level
performance on image net classification, 1026-1034, 2015.
PT

20. Keras Documentation. About Keras models. https://keras.io/models/about-keras-

models/. (Last accessed: 09 March 2017).
CE

21. S. Kiranyaz, T. Ince, M. Gabbouj. Real-time patient-specific ECG classification by 1-D

convolutional neural network. IEEE Transactions on Biomedical Engineering 63(3): 664-
675, 2016.
AC

22. A. Krizhevsky, I. Sutskever, G. E. Hinton. Imagenet classification with deep

convolutional neural networks. Advances in neural information processing systems: 12:
1097-1105, 2012.
23. Y. LeCun, L. Bottou, Y. Bengio, P Haffner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE 86 (11): 2278-2324, 1998.
 ACCEPTED MANUSCRIPT

24. Y. LeCun, Y. Bengio. Convolutional networks for images, speech, and time-series. In:
The handbook of brain theory and neural networks, MIT Press Cambridge, MA, USA,
1998.
25. R. J. Martis, U. R. Acharya, H. Adeli. Current methods in electrocardiogram
characterization. Computers in Biology and Medicine 48: 133-149, 2014.
26. R. J. Martis, U. R. Acharya, H. Adeli, H. Prasad, J. H. Tan, K. C. Chua, C. L. Too, S. W. J.
Yeo, L. Tong. Computer-aided diagnosis of atrial arrhythmia using dimensionality

T
reduction methods on transform domain representation. Biomedical Signal Processing

IP
and Control 13: 295-305, 2014.

CR
27. R. J. Martis, U. R. Acharya, H. Prasad, K. C. Chua, C. M. Lim, J. S. Suri. Application of
higher order statistics for atrial arrhythmia classification. Biomedical Signal Processing
and Control, 8: 888-900, 2013.

Assessing the costs of US

28. National Institute on Aging – turning discovery into health. Global health and aging.
aging and health
https://www.nia.nih.gov/research/publication/global-health-and-aging/assessing-costs-
care.
AN
aging-and-health-care. (Last accessed: 24 February 2017).
29. K. Najarian, R. Splinter. Biomedical signal and image processing, Second edition. CRC
M

Press, Taylor and Francis Group, Boca Raton, 2012.

30. M. Oquab, L. Bottou, I. Laptev, J. Sivic. Is object localization for free? - weakly-
supervised learning with convolutional neural networks. Proceedings of the IEEE
ED

Conference on Computer Vision and Pattern Recognition, 685-694, 2015.

31. B. N. Singh, A. Tiwari. Optimal selection of wavelet basis function applied to ECG signal
PT

denoising. Digital Signal Processing 16(3): 275-287, 2006.

32. K. Sirinukunwattana, S. E. A. Raza, Y. W. Tsang, D. R. J. Snead, I. A. Cree, N. M. Rajpoot.
CE

Locality sensitive deep learning for detection and classification of nuclei in routine colon
cancer histology images. IEEE Transactions on Medical Imaging 35(5): 1196-1206, 2016.
33. United Nations. Department of Economic and Social Affairs Population Division. World
AC

population aging 2015. New York, 2015.

34. Y. Wang, Y. S. Zhu, N. V. Thakor, Y. H. Xu. A short-time multifractal approach for
arrhythmia detection based on fuzzy neural network. IEEE Transactions on Biomedical
Engineering 48(9): 989-995, 2001.
 ACCEPTED MANUSCRIPT

35. Z. Yan, Y. Zhan, Z. Peng, S. Liao, Y. Shinagawa, S. Zhang, D. N. Metaxas, X. S. Zhou.

Multi-instance deep learning: discover discriminative local anatomies for bodypart
recognition. IEEE Transactions on Medical Imaging 35(5): 1332-1343, 2016.
36. M. Zubair, J. Kim, C. W. Yoon. An automated ECG beat classification system using
convolutional neural networks. IEEE 6th International Conference on IT Convergence
and Security, 2016.

T
IP
CR
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Graphical abstract

T
IP
CR
US
AN
M
ED
PT
CE
AC