SPECIAL SECTION ON DEEP LEARNING

FOR COMPUTER-AIDED MEDICAL DIAGNOSIS

Received April 26, 2019, accepted May 11, 2019, date of publication May 24, 2019, date of current version June 21, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2918792

A Novel Deep Arrhythmia-Diagnosis Network

for Atrial Fibrillation Classification Using
Electrocardiogram Signals
HAO DANG 1,2 , MUYI SUN 1,2 , GUANHONG ZHANG1,2 , XINGQUN QI 1,2 ,

XIAOGUANG ZHOU1,2 , AND QING CHANG3,4

1 School of Automation, Beijing University of Posts and Telecommunications, Beijing 100876, China
2 Engineering Research Center of Information Network, Ministry of Education, Beijing 100876, China
3 Shanghai General Practice Medical Education and Research Center, Shanghai 201800, China
4 Jiading District Central Hospital Affiliated Shanghai University of Medicine & Health Sciences, Shanghai 201800, China

Corresponding authors: Xiaoguang Zhou (zxg@bupt.edu.cn) and Qing Chang (robie0510@hotmail.com)

This work was supported in part by the Open Foundation of State Key Laboratory of Networking and Switching Technology, Beijing
University of Posts and Telecommunications, under Grant SKLNST-2018-1-18, and in part by the Foundation of Jiading Health Science
under Grant 2018-KY-04.

ABSTRACT Atrial fibrillation (AF), a common abnormal heartbeat rhythm, is a life-threatening recurrent
disease that affects older adults. Automatic classification is one of the most valuable topics in medical
sciences and bioinformatics, especially the detection of atrial fibrillation. However, it is difficult to accurately
explain the local characteristics of electrocardiogram (ECG) signals by manual analysis, due to their small
amplitude and short duration, coupled with the complexity and non-linearity. Hence, in this paper, we propose
a novel deep arrhythmia-diagnosis method, named deep CNN-BLSTM network model, to automatically
detect the AF heartbeats using the ECG signals. The model mainly consists of four convolution layers: two
BLSTM layers and two fully connected layers. The datasets of RR intervals (called set A) and heartbeat
sequences (P-QRS-T waves, called set B) are fed into the above-mentioned model. Most importantly, our
proposed approach achieved favorable performances with an accuracy of 99.94% and 98.63% in the training
and validation set of set A, respectively. In the testing set (unseen data sets), we obtained an accuracy
of 96.59%, a sensitivity of 99.93%, and a specificity of 97.03%. To the best of our knowledge, the algorithm
we proposed has shown excellent results compared to many state-of-art researches, which provides a new
solution for the AF automatic detection.

INDEX TERMS Convolution neural network, deep learning, bi-directional long short-term memory, atrial
fibrillation.

I. INTRODUCTION its incidence increases with age [1], [2]. AF is characterized

According to the report of World Health Statistic 2017 from by rapid and irregular ventricular contractions [3], which is
the World Health Organization (WHO), non-communicable divided into paroxysmal atrial fibrillation, persistent atrial
diseases accounted for 70% of the total deaths, and fibrillation, and permanent atrial fibrillation [4].
the most primary causes of death were cardiovascular The abnormal heartbeat rhythm is related to the high risk
disease (CVD) [1]. The number of patients died of cardio- of stroke, heart disease, and cardiovascular disease [2], [5].
vascular disease accounted for more than 40% of deaths In addition, there is also a close relationship between AF
from residents, ranking first comparing to cancer and other and obesity [6], hypertension mitral or tricuspid valvular
diseases [1]. Thereinto, atrial fibrillation (AF) is the most disorders, obstructive sleep apnea [7], hyperthyroidism [8],
common clinically sustained cardiac arrhythmia disorder, and alcoholism [9]–[11], which can not only lower the quality
with a prevalence of approximately 1% in the population, and of life, but also increase the risk of mortality. Accordingly,
AF is the most common arrhythmia disease, and it is neces-
The associate editor coordinating the review of this manuscript and sary to automatically detect AF for early diagnosis, manage-
approving it for publication was Shuihua Wang. ment, and prevention of related complications [12].
2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 7, 2019 Personal use is also permitted, but republication/redistribution requires IEEE permission. 75577
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
At present, the diagnosis of AF mainly depends on the
presence of some typical symptoms (such as breathing dif-
ficulties, chest pain, abnormal pulse, and so on.) [12] and the
characteristics of electrocardiogram (ECG). However, it is
difficult to accurately detect AF in the early stage, mainly
because sometimes there are no obvious symptoms when
AF occurs [13], and well-trained professional physicians
are required to accurately determine the feature information
of ECG [12], [14]. Therefore, it is important and valuable
to develop a fast and accurate algorithm of AF automatic
detection.
In recent years, deep convolutional neural networks
(DCNN) are widely applied to analyze images, such an image
classification, objection detection, and image segmentation.
With the development of DCNN, deep networks naturally
provide a new research idea for different research directions
due to different network structures, depth, and width, which
can integrate low/mid/high-level features [15] and classi-
fiers in an end-to-end multilayer model. Simultaneously, the
‘‘levels’’ of features can be enriched by the number of stacked
layers [16]. However, conventional ECG recognition sys-
tems often rely on well-trained professional physicians and
also require complex feature extraction processes, increasing
the computational time to classify arrhythmias. This makes
them less reliable in the clinic or practice, as their accu-
racy and efficiency often vary greatly for larger databases,
which also is a motivation of CNN-BLSTM for ECG
classification.
In our study, a reliable arrhythmia-diagnosis network based
on deep learning for the automatic detection of AF is pro-
posed. The network includes three main stages of data
processing, feature extraction, and classification. Initially,
the peaks detection and data segmentation are required to
obtain the input data, which is fed into the second stage.
Secondly, the automatic feature extraction is implemented by
four convolution layers and two BLSTM blocks. Then the
ECG signal episodes are classified by two fully-connected
layers and a SoftMax layer. Ultimately, we quantitatively
investigate the performance of the proposed network and
the function of different tricks (such as batch normaliza-
tion, different epoch, and so on.) by experiments. In general,
the proposed method does not need the traditional hand-
crafted feature extraction methods, and shows state-of-the-art
results. It provides a new idea for the classification of
one-dimensional signals.
II. RELATED WORK
In this section, we discuss in detail the approaches that
closely related to this work: machine learning and deep
learning approaches, which are widely used to detect AF.
However, due to the limitations of space and capacity, we can-
not provide a complete and comprehensive research report,
although we have collected a large number of state-of-the-art
researches.
75578
A. MACHINE LEARNING
Since the last decades, the pattern recognition methods based
on machine learning has been widely researched and applied
for AF detection and classification [17]–[21]. For instance,
Liu et al. [17] proposed a normalized fuzzy entropy, a novel
entropy measure suitable to enhance the performance of
entropy-based AF detectors using a short-term RR interval.
In their work, the detector was trained using the MIT-BIH
atrial fibrillation databases (MIT-BIH AF), and tested on
the MIT-BIH normal sinus rhythm (MIT-BIH NSR) and
MIT-BIH arrhythmia databases (MITAB), with the high-
est area under receiver (AUC) values of 92.72%, 95.27%,
and 96.76% for 12−, 30−, and 60-beat window lengths
respectively. Lake and Moorman et al. [18] developed a
new method of entropy estimation in short time series and
implemented them to detect AF. The proposed algorithm
has achieved a high degree of accuracy in distinguish-
ing AF from normal sinus rhythm in 12-beat. An average
sensitivity of 91% and specificity of 98% were obtained.
Liu et al. [19] used an SVM-based method to detect the
four arrhythmias included AF, which achieve a moderate
performance with an accuracy of 86.23% and 87.71% on
the training and validation set of PhysioNet/Computing in
Cardiology Challenge 2017. Zhao et al. [20] also developed
a new entropy measure, named EntropyAF , to detect AF,
which was trained using the MIT-BIH AF, and tested on
the clinical wearable long-term AF recordings. EntropyAF
acquired a performance with an accuracy of 87.10%, a sensi-
tivity of 92.77%, and a specificity of 85.17%. Mei et al. [21]
presented a method based on heart rate variability and spectral
features to detect AF from short single lead ECG recording.
For two classes classification problem (Normal and AF),
accuracy varies from 92.0% to 96.0% under different addi-
tional noise levels. Excellent accuracy and very high speci-
ficity of this model made it a competitive algorithm for
preliminary screening.
Otherwise, we review the special issue about AF detec-
tion from Physiological Measurement journal [22]–[34].
The review mainly summarizes the three parts included the
source of data, features selection, and popular algorithm.
The data used in [22]–[34] mainly includes three aspects:
common public databases (such as MITAB, MIT-BIH AF,
wrist-worn devices, and the data from PhysioNet/CinC Chal-
lenge dataset and so on.). The four chief features, such as
AF features, RR interval features, morphology features, and
features of similarity index between beats, are selected to
detect AF signals.
In general, some machine learning classifiers included
support vector machining (SVM), AdaBoosted Decision
Tree (ADT), Artificial Neural Network (ANN), Gradient
Boost Decision Tree (GBDT), Coefficient of Sample Entropy
(CosEn), and Linear Discriminant Analysis (LDA) are espe-
cially popular in the above researches. Those methods can
be divided into five steps: feature perception, data pre-
processing, features extraction, feature selection, signals
VOLUME 7, 2019
H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
prediction, and classification. The classification of the
ECG signal has witnessed a quantum leap in the performance
on benchmark datasets.
Nevertheless, with the development of machine learn-
ing, conventional models have several insurmountable draw-
backs. First, most of the reported approaches in the literatures
for classifying the ECG signals solely rely on the traditional
methods of computer vision with machine learning, which
separate feature extraction from classifier design, then com-
bine them when applied. Second, handcrafted extract features
methods are widely used in the traditional machine learning
pipelines for the ECG signals classification. And it mainly
depends on a large number of expert priori knowledge and
sufficient capacity of biomedical signal processing. On the
basis of this, it is also necessary to design a good classifier
algorithm, but is difficult to achieve the optimal effect. So, it is
important to acquire representative, dominant, and critical
features for a particular ECG signal and simultaneously
design excellent and reliable classifier. In addition, it is
time-consuming and strenuous to analyze the ECG signals,
due to the effect of computer hardware. Moreover, because
of the unique characteristics of ECG signals: small
amplitude (mV) and small duration (sec), it also is a challenge
to analyze the ECG signals. Hence, the interpretation of these
long duration of signals may lead to inter and intra-observer
variabilities [35].
B. DEEP LEARNING
Recent years, deep learning regarded convolution neural net-
works (CNN) as the core has made a series of breakthroughs
for image classification, objection detection, and semantic
segmentation. The deep learning has become a dramatical
method to solve the research problem of computer vision, due
to the advances in the CNN model, GPU hardware and the
feature of large-scale data.
Therefore, it appears natural to leverage convolu-
tional layer to extract features of ECG signals and
solve the problems of machine learning mentioned above.
Acharya et al. [36] implemented a CNN model to detect
a normal and myocardial infarction(MI) ECG beats, which
achieved an average accuracy of 93.53% and 95.22% based
ECG signals. Al Rahhal et al. [37] proposed deep learn-
ing approach for active classification of ECG signals and
obtained significant accuracy improvements with less expert
interaction and faster online retraining compared to the
state-of-the-art methods in 2016. Acharya et al. [38] also
presented a CNN technique to automatically detect the dif-
ferent ECG segments included the normal, atrial fibrilla-
tion, atrial flutter, and ventricular fibrillation. They have
acquired an accuracy of 94.90%, a sensitivity of 99.13%, and
a specificity of 81.44% for five seconds of ECG duration.
Pourbabaee et al. [39] developed a deep learning machine to
screen and identify patients with paroxysmal atrial fibrillation
(PAF). The extensive simulations indicated that combining
the learned features with other classifiers will significantly
improve the performance of the patient screening system.
VOLUME 7, 2019
Xu et al. [40] proposed a new model that combines modified
frequency slice wavelet transform (MFSWT) and convolu-
tional neural network (CNN) for automatic AF beat detection.
The method was used on the MIT-BIH AF and achieved
a mean accuracy of 81.07% from 5-fold cross-validation.
The corresponding sensitivity and specificity performances
were 74.96% and 86.41%. Andersen et al. [41] developed an
end-to-end model combining the CNN and recurrent neural
network (RNN) to classify the AF and normal sinus rhythm
(NSR). It showed an accuracy, a sensitivity, and a specificity
of 97.8%, 98.98%, and 96.95%, validated through 5-fold
cross-validation. And it also achieved the results of accuracy
(87.40%), sensitivity (98.96%), and specificity (86.04%) on
unseen data sets.
However, it remains a challenging work to build an effec-
tive learning idea for ECG signals, due to itself specificity
features of small datasets and one-dimensional signal. Mean-
while, it is also very difficult to improve the detection speed
and accuracy.
In our paper, we propose a novel deep arrhythmia-
diagnosis method, namely deep CNN-BLSTM network,
to automatically detect AF from ECG signals. In summary,
the contributions of this work are as follows:
(1) We propose an end-to-end deep CNN-BLSTM network
for the detection of AF signals due to its features of low
dimensionality and small memory, which replaces the hand-
crafted features with CNN and BLSTM.
(2) We design multi-scale signals included RR intervals
(named set A) and Heartbeat sequences (P-QRS-T waves,
named set B.) as the inputs of the network. This design can
be leveraged to extract multi-scale features, and then improve
the generalization of the network model.
(3) We present a detailed experimental study of ECG signal
classification with significantly improved performance.
(4) Last, we show that our proposed CNN-BLSTM net-
work achieves state-of-the-art results on datasets and carries
out the full comparisons with other methods.
III. DESCRIPTION AND PRE-PROCESSING OF DATASET
In this section, on one hand, we mainly introduce the database
used in this paper, which includes the basic information
and the chief characteristics of the AF signal. On the other
hand, the pre-processing of dataset, such as peaks detection,
segmentation, and normalization, is described in this work.
A. DESCRIPTION OF DATASET
In this paper, the ECG signals are acquired from the
MIT-BIH Atrial Fibrillation databases (MIT-BIH AF) [42]
from physio net. The database includes 25 long-term ECG
records (mostly paroxysmal atrial fibrillation). Of these,
23 records contain two ECG signals (in the.dat file). Each sig-
nal record lasts for 10 hours with a sampling rate of 250 and
a resolution of 12 bits. A typical normal ECG signal of
a heartbeat is shown in Fig.1, which includes P-wave,
Q-wave, R-wave, S-wave, QRS-complex, T-wave, U-wave,
PR-interval, PR- segment, ST-segment, and QT-segment.
75579
 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
FIGURE 1. A typical normal ECG signal of a heartbeat (Source: Ref. [43]).
A normal and integrated heartbeat is comprised of P-wave, PR-interval,
Q-wave, R- wave, QRS-complex, S- wave, T- wave, ST-segment,
QT-segment, and U- wave. Each waveform corresponds to the
physiological process of cardiac excitation. The total
duration of a heartbeat is about 0.8s.
FIGURE 2. AF signal segments of different samples. It is observed that
these segments represent AF signal feature though the time periods are
different. The characteristic information of AF contained in each signal is
different.
Otherwise, the database mainly includes four kinds of heart-
beat annotations: atrial fibrillation, atrial flutter, AV junc-
tional rhythm, and all other rhythms. Fig.2 shows AF signal
segments from different time periods, such as 1s and 3s.
Atrial fibrillation (AF), a type of abnormal heart rhythm,
is characterized by a rapid and irregular rhythm caused a
series of threats to the human body. Specifically, in four
aspects:
(1) P waves disappear, which replace with F-wave of dif-
ferent sizes and shapes;
(2) The F-wave frequency is 350 − 600 beats/min, and its
size, shape, and amplitude are different. The P-R interval is
not easy to determine;
(3) RR interval is absolutely uneven, and ventricular rate
generally increases, but usually <160 beats / min;
(4) There is a normal QRS-complex when there is no differ-
ence in terms of intraventricular conduction, and when intra-
ventricular differential conduction occurs, the QRS-complex
is broadly deformed.
In this paper, we aim at researching the classification of
AF rhythms. So, we only consider whether the signal seg-
mentation is the AF or not.
75580
FIGURE 3. The detection of R peak. The black dots represent the position
of R peak of the ECG signal. The RR intervals is an important feature for
the detection of AF.
B. PRE-PROCESSING OF DATASET
In our work, the ECG signals from the public database are
pre-processed as follows:
(1) Peak detection. This stage is to determine the positions
of all interested peaks in the ECG signals. The detection of
R-peak is essential for set A. And the detection of P-wave,
QRS-complex, and T-wave contribute to the construction of
set B. Fig.3 gives the detection of R peak.
(2) The setting of two datasets. In our paper, we divide
the AF database into two data sets according to different
input signals. According to the RR interval, the ECG sig-
nals are extracted base on the detection of R-peak, which is
called set A. On the other hand, the P-wave, QRS-complex,
and T-wave detected in step (1) are divided into a heartbeat
sequences centered around R-peak, which are called set B.
Fig.4 shows the samples of set A and set B.
(3) The segmentation of ECG signals. On the basis of
above, the integrated RR intervals of a dataset are divided
into a mass of overlapping sequences included 100 samples
(or 100 values, not 100 beats, it is 99 samples overlap).
A RR interval of set A only contains 100 values, which will be
input the model. Similarly, an integrated heartbeat sequence
(shown in Fig. 1) is from P onset to T offset (or U onset). The
heartbeat sequences of set B is divided into a segment, which
contains 500 values centered around R-peak, which also will
be input the model. According to the label information of AF
in the annotation file(.atr), all the input data segments are
labeled. As long as a RR interval (set A) or one heartbeat
(set B) is labeled as AF, the label of this data segment is
considered to be AF, and the labels of all other data segments
are normal signals without AF features.
(4) Z-score normalization. In order to address the prob-
lem of non-standard data and improve the comparability of
data, the new set A and set B are normalized by z-score
normalization.
IV. NETWORK ARCHITECTURE
In this part, we describe the network structure model pro-
posed in this study, which mainly includes four convolu-
tion layers, two BLSTM layers, two full connection layers,
and other supporting computing operations (pooling layers,
ReLU activate, batch normalization, dropout, and so on.).
VOLUME 7, 2019
H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
FIGURE 4. Set A: RR intervals (left) and set B: Heartbeat sequences
(right). In the MIT-BIT AF database, each dataset has a duration
of 10 hours, which include a number of RR intervals. The integrated RR
intervals of a dataset are divided into a mass of overlapping sequences
included 100 samples (or 100 values, not 100 beats, it is 99 samples
overlap). The RR intervals of set A only contain 100 values, which will be
input the model. Similarly, each dataset also includes a mass of
heartbeats (An integrated heartbeat is shown in Fig. 1.).
An integrated heartbeat is from P onset to T offset (or U onset). The
heartbeat sequences of set B only contain 500 values of P, Q, R, S, T,
which also will be input the model.
We first review the network model of CNN and bi-directional
long-term memory network (BLSTM), which are closely
related to the model structure proposed in this paper. Then,
the most important is put forward the research model of this
paper, and describe the structure, parameters, and mathemat-
ical expression of the model in detail.
A. CONVOLUTIONAL NEURAL NETWORK
Since the CNN initially was proposed in the 1980s by
Fukushima [44], it has been spectacular advances in deep
learning. Generally, the CNN network model is processing of
feature extraction and learning, which mainly includes input
layer, convolution layer, polling layer, activation function
layer, fully connected layer or fully convolution layer, and
SoftMax layer.
1) CONVOLUTION LAYER
The essence of convolution layer is to perform corresponding
mathematical calculation on vectors, which is affected by the
shape of its input as well as the size of the kernel, stride,
and padding. It is applicable to any input type of included
an image, a text or a sound slip. The chief function is to
extract feature map on the given input signal. In this paper,
VOLUME 7, 2019
it is applied to analyze ECG signals, and its main function is
to carry on the process of features learning.
2) POOLING LAYER
In essence, the pooling layer is to retain maximum spatial
information while reducing the number of feature maps by
operating corresponding mathematical calculation, such as
max pooling or average pooling.
3) ACTIVATION FUNCTION
Activation function is used to introduce nonlinearity in the
values of the incoming layer, so that the neural network can
solve more complex problems, but also enhance the feature
expression and learning ability of the neural network model.
In this paper, the ReLU is regarded as the activation function
in convolutional layers, while the Sigmoid and tanh are used
as the activation function in BLSTM blocks.
4) FULLY CONNECTED LAYER
It plays a role of a classifier in the CNN, mapping the dis-
tributed feature representation learned from the convolutional
layer.
5) SOFTMAX LAYER
SoftMax function maps multiple scalars to a probability dis-
tribution with each value output in the range (0,1), which is
often used in the last layer of the neural network as the output
layer for multi-classification.
In this paper, the input ECG signal samples are first fed
into 3 convolution layers and 3 max-pooling layers with a
stride of 1 and 2 respectively. The main function is to extract
effective features from input samples.
B. BI-DIRECTIONAL LONG SHORT-TERM MEMORY
NETWORK
As an important branch of deep learning, Recurrent Neural
Networks (RNN) [45] has always been the research focus.
RNN introduces the concept of loop, which is mainly aimed
at solving problems related to time series signals (such as
continuous sound signals or continuous handwritten words.).
RNN takes advantage of itself internal memory to process
input sequences in any duration, which makes it easier to
deal with such problems as unsegmented handwriting recog-
nition or speech recognition.
The huge advantage of RNN is to learn the information
characteristics of different time effectively. However, it has
appeared that actual initial information lost over time, which
is called a long-term dependence problem. So, the Long-
short-term-memory [46], [47] (LSTM) network arises at the
historic moment, which can learn characteristics of the previ-
ous stages from input signals. Although the LSTM memory
unit can deal with uncertain long time series data, the tra-
ditional model of LSTM often ignored the information in
the future stage, only processing the data along the direc-
tion of forward. The future input information can also pro-
vide insight about hidden states that can help prediction.
75581
 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
FIGURE 5. The Bi–directional long short-term memory (BLSTM) network. The BLSTM allows the output units yi to
automatically learn a representation that depends on both the previous and the future information.
And it is sensitive to the input values around time t.
In recent years, a novel method based on the bidirectional
long short-term memory (BLSTM) be proposed, which is uti-
lized to extract the bidirectional information from the forward
model (operating from t0 , t1 , . . . , to tn .) and backward model
(operating from tn , tn−1 , . . . , to t0 .) at the same time.
BLSTM is a recurrent neural network that has an input
layer, two hidden layers, and an output layer. The hid-
den layers are a memory block included three gate mecha-
nisms called input gate, forget gate, and output gate. These
gate mechanisms control the amount of information and
the extraction of features. The input gate determines what
new information will be stored in the cell state, the forget
gate determines what information will be retained or dis-
carded from the cell state. After updating the cell state,
the output gate completes the output of network model.
Recently, BLSTM model has been widely researched in
the continuous time series signals fields such as speech
recognition.
In our paper, an important BLSTM block is leveraged
in the network architecture. As illustrated in Fig.5, a stan-
dard BLSTM block includes an input layer, a forward layer,
a backward layer, and an output layer, which realizes the
function processing the data in both directions of two sep-
arate hidden layers and then merging to the same output
layer. The learning of both directions is able to obtain a
more accurate prediction result. In Fig.5, a BLSTM block
computes the forward hidden layer Fi , the backward hidden
layer Bi , and the output layer yi by updating the following
equations (1)-(3):
Fi = f (w1 xi + w2 Fi−1 + bi ) (1)
Bi = f (w3 xi + w5 Bi1 + bj ) (2)
yi = g(w4 Fi + w6 Bi−1 + bk ) (3)
where, f is the cell state vectors update function in the hidden
layers of the BLSTM block, which is detailly defined by
75582
composite equation (4) to (9).
ft = σ (wf [ht−1 , xt ] + bf ) (4)
it = σ (wi [ht−1 , xt ] + bi ) (5)
jt = tanh(wc [ht−1 , xt ] + bj ) (6)
ot = σ (wo [ht−1 , xt ] + bk ) (7)
ct = ft ∗ct−1 + it ∗ jt (8)
ht = tanh(c)t ∗ kt (9)
where ft , it , ot , and ct respectively represent the forget gate,
input gate, output gate, and cell state in long short-term
memory (LSTM) network (Fig.6), jt is a tanh layer created
a new candidate value vector, which is added to the state. The
cell state is processed through the tanh (to get a value between
−1 and 1) and multiplied by the output of the sigmoid gate.
Then we obtain the final output. The σ and tanh are a usual
element-wise sigmoid or tanh activation function. The sym-
bols of ‘‘⊕’’ and ‘‘⊗’’ denote the element-wise concatenated
and multiplied.
C. THE PROPOSED ARCHITECTURE
In this study, a novel deep CNN-BLSTM network
(Fig.7) model included four convolution layers and two
BLSTM blocks is proposed to classify the AF signals.
Generally, the CNN model has been widely implemented
to analyze two-dimensional images signals in many litera-
tures. However, in our paper, the CNN network model is
applied to process the one-dimensional ECG signals. The
proposed architecture is shown in Fig.7 and detailed param-
eters are listed in Table 1. The Fig.8 shows the operation
details between two BLSTM blocks, two fully-connected
layers, and SoftMax layer. Detailly, the deep CNN-BLSTM
network is composed of four convolutional layers, two
BLSTM blocks, two fully-connected layers, a SoftMax layer,
and corresponding activate function, batch normalization,
global max-pooling, and dropout layer. Most importantly, the
VOLUME 7, 2019
H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals

FIGURE 6. The LSTM network. The σ and tanh represent an element-wise sigmoid and tanh activation function. The symbols of ‘‘⊕’’ and ‘‘⊗’’ denote the
element-wise concatenation and multiplication.

FIGURE 7. The proposed novel deep CNN-BLSTM network in this paper. FM represents feature map; MP denotes max-pooling. The operations of four
convolutional layers and two BLSTM blocks show the processing of feature learning. The operation of concatenate is a layer that connects multiple inputs
to one output that implements the joint of input data.

network architecture includes two different input signals.
Set A is the overlapping sequences of RR intervals and set
B is the Heartbeat sequences (P-QRS-T waves).
V. EXPERIMENTAL RESULTS
In order to fully analyze and verify the effectiveness of the
model proposed in this paper, we train and validate the two
data sets of set A and set B described in the III section respec-
tively according to the design of the model, and then fully
comparing the different results by adjusting key parameters.
We first describe the main initialization parameters, opti-
mization function in the model training stage. Furthermore,
three evaluation metrics including accuracy, sensitivity, and
specificity are calculated. Finally, we make a comprehensive
analysis of the performance metrics and compare with impor-
tant references in recent years.
A. IMPLEMENTATION DETAILS
In the standard deep CNN-BLSTM network, the first three
convolution layers (from stage one of Conv to stage three
of Conv.) are followed by the ReLU activate function, which
increases the nonlinear representation ability of network
model. The activate function is sigmoid and tanh in the
BLSTM blocks. Then the last convolutional layer is used
to adjust the output from BLSTM block, which is bene-
ficial to the operation of global max-pooling. Meanwhile,
it also can adequately learn features from the signal. Global
max-pooling is implemented between the BLSTM block

VOLUME 7, 2019 75583

 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals

TABLE 1. The detailed parameters of network architecture.

and the first fully-connected layer to compress the features

of the output sequences from BLSTM block. The fully-
connected layers are followed by batch normalization, ReLU,
and dropout. The network model is carried out by optimizing
the cross-entropy function with Adam optimizer [48], which
utilizes the TensorFlow [49] on a Titian Xp GPU with batch FIGURE 8. The details of two BLSTM blocks.
size 1024 for 100 epochs.
The most important calculation of CNN model is gradient
The cross-entropy loss function of two class problem is
backpropagation and weight update. The weights and biases
obtained according to equation (12).
are updated according to equation (10), (11).
L yt , yp =−(yt ∗ log yp +(1−yt )log(1−yp )yt ∈{0, 1} (12)

∂c
 
rλ where, yt represents the ground truth of each sample, yp is a
wl = 1 − w−r (10)
n ∂w probability predicted by the training model.
∂c Otherwise, the two datasets are divided into three phases
bl = bl−1 − r (11)
∂w of training (80%), validation (10%), and testing (10%). The
training set is used to train the model and determine the model
where w, r, λ,n and c represent the weight, learning rate, parameters. The validation set is used to tune parameters and
regularization parameter, the number of training samples and determine the optimal model. And the test set, which is not
the cost function respectively. been used before, is utilized to evaluate the performance of
According to the earlier experiments, the initial learning the trained model. The detailed results will be shown in part
rate and decay rate are set to 0.001 and 0.9, respectively. The of section B.
cross-entropy function is leveraged to evaluate the loss of the In addition, a 10-fold cross-validation strategy is employed
model. During the stage of fully-connected layer, the dropout in our study. The set A and set B are randomly divided
is adopted to reduce overfitting and improve the generaliza- into 10 equal segments after the generation of signal seg-
tion. The parameter of dropout is set to 0.2 taking into account ments.9/10 parts are used to training and 1/10 parts are used
the data difference and the number of neurons. to validation of the proposed model. The approach is repeated

75584 VOLUME 7, 2019

H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals

10 times by iterating all the data. Then, the evaluation metrics TABLE 2. The overall classification results for the AF class.
of accuracy, sensitivity, and specificity are shown in each
step. Finally, the overall evaluation metrics is achieved by
averaging the results of the 10 steps.

B. PERFORMANCE EVALUATION
In order to evaluate the performance of the proposed network
architecture, we assume the class containing AF signals as
the positive class, and therefore the accuracy, sensitivity, and
specificity are computed, respectively.
TP + TN
Accuracy = (13)
TP + FP + TN + FN
TP
Sensitivity = (14)
TP + FN
TN
Specitifity = (15)
TN + FP
TP means true positive, which is the number of AF signals
classified correctly;
FP means false positive, which is the number of AF signals
classified wrongly;
TN means true negative, which is the number of signals
without AF classified correctly;
FN means false negative, which is the number of signals
without AF classified wrongly.
In the experiment, in addition to a standard network model
(shown in Table 1), we set up three comparison experiment
(named 1, 2, 3.) through parameter modulation, which is
FIGURE 9. Training and validation accuracy with 100 epochs in set B.
aimed at verifying the superiority of the model. The dif-
ference between experiment 1 and the standard model is
that epoch is decreased to 80. And the BLSTM block in
the model is replaced with the conventional LSTM block in
experiment 2. And in experiment 3, the BN layer between the
last fully connected layer and output layer is removed, while
the BN layer also is removed among the two full-connected
layers. Fig.9 and Fig.10 show the accuracy rate and the loss
of training and validation set with 100 epochs when the input
is set A.
As can be seen from the Fig.9 and Fig.10, the performance
of deep CNN-BLSTM network on the training set is slightly
better than the validation set. Meanwhile, the change of loss
FIGURE 10. Training and validation loss with 100 epochs in set A.
illustrates that the model converges to a steady value and there
is no over-fitting phenomenon. The maximum accuracy of
training set is 99.92%, and maximum accuracy is 98.94% in
the validation set. AF is closely related to the feature information of RR interval.
When the performance of the model is measured in terms So, the result of set A is better than set B. Therefore, we aban-
of time costs, the average epoch last 192 seconds using don the data of set B in the following experiment. Otherwise,
a Titian Xp GPU during the training stage. In the testing compared with experiment 1 (80 epochs), Fig.13 shows the
stage, we cost 54 seconds to 104,858 signal records. The accuracy with 100 and 80 epochs in the training and valida-
average signal cost is about 0.51ms, which leads to real-time tion set. It can be seen from the Fig.13 that the change trend is
speed theoretically. The average accuracy, sensitivity, and almost same, but the accuracy variation trend is more stable
sensitivity for the testing set are given in Table2. Meanwhile, than that of 80 epochs during the stage of BLSTM network
set B is also used as input for training. Fig.11 and Fig.12 show model with 100 epochs. Meanwhile, Fig. 14 shows that the
the accuracy is worse than the result of set A, although the loss BLSTM model has better performance compared with LSTM
is also convergent in the end. The accuracy is less than 85% network under the condition of consistent parameters, which
in the training set and validation set. The main reason is that fully reflects the superiority of BLSTM model.

VOLUME 7, 2019 75585

 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
FIGURE 11. Training and validation accuracy with 100 epochs in set A.
FIGURE 12. Training and validation loss with100 epochs in set B.
FIGURE 13. The accuracy with 80 and100 epochs.
Otherwise, the chief function of Batch Normalization (BN)
is to make the input of each layer maintain the same distri-
bution in the process of deep neural network training, so as
to eventually alleviate the gradient disappearance/explosion
phenomenon and accelerate the training speed of the model.
By comparing the result of experiment 3 (Fig.15), it can
be concluded that the operation of BN layer has certain
effect, but the difference is relatively weak. In our experi-
ment, the change trend with BN (blue solid) and that trend
without BN (yellow solid) in the training set is almost same
in Fig.15. And the average accuracy of standard network
model and experiment 3 differed by 0.3% in the valida-
tion set. The main reason is that the network model for
ECG signal research is relatively simple compared with the
image classification model, and the main role of BN is rel-
atively weak. Finally, all the experimental results are shown
in Table 2.
75586
FIGURE 14. The accuracy of BLSTM and LSTM model.
FIGURE 15. The accuracy with BN and without BN.
C. COMPARISON AND DISCUSSION
Table 3 summarizes the various techniques used by the
researchers to automatically detect AF using ECG signals.
It mainly includes research literatures, databases, meth-
ods, and three evaluation metrics of accuracy, sensitivity,
and specificity. The chief methods contain morphological
analysis, machine learning, and popular deep learning in
recent years. These studies have achieved good results.
Zhou et al. [10] used a novel recursive method to detect
AF episodes, which achieved a test performance of 96.89%,
98.25%, and 97.67% with respect to the sensitivity, speci-
ficity, and overall accuracy. Although the two performances
of specificity and accuracy are slightly higher than the result
of our method, Zhou [10] method involves several basic
operations of nonlinear filters, symbolic dynamics, and the
calculation of Shannon entropy, which will consume a lot of
time with increasing computation. Petrėnas et al. [50] utilized
ectopic beat filtering and bigeminal suppression for detec-
tion of subclinical AF episodes. Asgari et al. [12] leveraged
the stationary wavelet transform and support vector machine
to detect AF episodes. Henzel et al. [51] demonstrated the
effectiveness of machine learning based on generalized linear
model evaluated with ROC. These three methods (Proposed
in [12], [50] and [51].) are same with the idea proposed
in [10], which belong to machine learning methods. Their
comprehensive performances are worse than the method we
proposed. Moreover, the main problem of machine learning
is that handcrafted extract features methods rely on a large
number of expert priori knowledge and sufficient capacity of
VOLUME 7, 2019
H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
TABLE 3. Summary of selected studies conducted for the automated detection of af heartbeats.
biomedical signal processing, which requires a lot of time and
computation.
In the MIT-BIH AF, the [17] method achieved the
sensitivity of 96.58%,96.71%, and 98.46% using 12-beat,
30-beat, and 60-beat RR segments respectively, the speci-
ficity of 83.31%, 87.52%, and 89.85% for the three-
time window types, the accuracy of 88.96%, 91.43%,
93.51%. And the [18] method based-COSEn also acquired
the results of sensitivity (91%) and specificity (98%) for
MIT-BIH AF, respectively. Compared with [17] and [18],
our study employs 100 samples (or 100 values, not 100-
beat) from RR intervals as the input signal, shows the
higher performances in MIT-BIH AF than [17] and [18]
methods.
Otherwise, we also evaluated the method based on
SVM proposed in [19] with respect to the new data from the
PhysioNet/Computing in Cardiology Challenge 2017. It is a
study of our team in Cardiology Challenge 2017. The results
showed the best accuracy is 86.23% and 87.71% on the train-
ing set and validation set with respect to all the feature sets,
which ranked the fourth in AF detection. However, the perfor-
mances are worse than the accuracy of 99.94% and 98.63%
VOLUME 7, 2019
on the training set and validation set in this paper. Since the
conventional feature extraction method usually takes most of
the time, it increases the computation compared with the deep
learning idea.
In view of the deep learning approaches are widely
employed to deal with AF, Acharya et al. [38] proposed a
multi-layer DCNN to automatically detect the AF segments.
They achieved the results of accuracy (92.50%, 94.90%), sen-
sitivity (98.09%, 99.13%), and specificity (93.13%, 81.44%)
for two and five seconds of ECG segments. Compar-
ing with the result of our study, the method mentioned
above is highlighted deficiency. For a detailed analysis,
Andersen et al. [41] showed a method that combining the
CNN and LSTM to classify the AF and normal sinus rhythm
(NSR). It showed an accuracy, sensitivity, and specificity
of 97.8%, 98.98%, and 96.95%, validated through 5-fold
cross-validation. And it also achieved the results of accuracy
(87.40%), sensitivity (98.96%), and specificity (86.04%) on
unseen data sets. This paper is very similar to our study.
Compared with the method we proposed, there are mainly
four differences: the structure of model, the input of signal,
the validation method, and the performance results.
75587
 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
First, the model proposed in our study includes four con-
volutional layers, two BLSTM blocks, two fully connected
layers, and other tricks (such as pooling layers, ReLU acti-
vate, batch normalization, global max-pooling, dropout, and
so on. It is shown in Fig. 7). The different from [41] is
that three convolution layers followed max-pooling layer.
And global max-pooling is leveraged between the second
BLSTM blocks and the fully connected layers to compress
the features from the BLSTM blocks. Otherwise, correspond-
ing tricks of batch normalization and dropout are used to
alleviate the gradient disappearance/explosion phenomenon
and overfitting of the model. By contrast, our method is
more complicated and comprehensive. The four convolu-
tional layers and two BLSTM blocks can adequately learn
features from the signal. Second, in [41], the raw ECG record-
ings from the databases are converted to RRI sequences.
The RRI sequences are fed into the model. In our study,
we design multi-scale signals included RR intervals (named
set A) and Heartbeat sequences (named set B.) as the inputs
of network. This demonstrates the generalization of the
model.
In addition, we obtain 1,038,946 records by segmenting the
ECG signals. Each record includes 100 RR intervals. Total
RR intervals is 103,894,600 records. This amount of data is
much larger than [41]. This is the main reason why I design
more complex model. Respecting to a large amount of data
and assessing the uncertainty of the performance, 10-fold
cross validation is leveraged in model validation.
Most importantly, our proposed approach achieved favor-
able performances in set A with an accuracy of 99.94%,
98.63% in training and validation set, respectively. In unseen
data sets, we obtained an accuracy of 96.59%, a sensitivity
of 99.93%, and a specificity of 97.03%, which is 9.19%,
0.97%, and 10.99% higher compared with [41], respec-
tively. This also proves that our method has better effect and
generalization.
In general, compared with the traditional morphology and
machine learning methods, our proposed method does not
consume a large amount of time and energy to selecting and
extracting features, all available information in the training
dataset is fed into the deep CNN-BLSTM network. Therefore,
it can extract the implicit characteristic information, so that it
can acquire state-of-art performances in an unseen dataset.
The average signal record cost is about 0.51ms, which leads
to real-time speed theoretically.
Meanwhile, we innovate the input signals of model and
network structure compared with the other deep learning
methods, and achieving a commendable effect on the training
set, validation set, and testing set.
A limitation of the present work is that the pro-
posed network is not evaluated on another public database
except MIT-AF. Otherwise, although we carry out training,
validation, and testing and achieve excellent performance
metrics on the public clinical database of MIT-AF, and it also
provides a new solution of AF detection. the model is not been
employed in the genuine clinical diagnosis and application
75588
of hospital. So, genuine clinical contributions need to be
strengthened in the future study.
Another limitation is that in order to reduce the complexity
of training, our input signals are only two points of RR peaks
between all the RR intervals, which do not involve the other
signal values between the RR intervals. Hence, the possibility
of further enhancement of the method using all the points of
RR intervals to be investigated.
VI. CONCLUSION AND FUTURE WORK
Automatic classification is the most valuable topics in medi-
cal and bioinformatics, especially the detection of AF. In this
paper, we have proposed a novel deep CNN-BLSTM network
to detect the AF signal from ECG records. The model inte-
grates the feature extraction methods of CNN and BLSTM,
which is composed of four CNN layers, two BLSTM blocks
and two fully connected layers. These CNN layers and
BLSTM layers are used to extract features of ECG signals.
Otherwise, two datasets of RR intervals (named set A) and
Heartbeat sequences (named set B.) mentioned in section III
are trained respectively. The performance of the set A is
largely higher than the set B. According to the analysis,
the main reason is that the signal of AF is closely related
to the information of RR intervals. Therefore, we abandoned
the data of set B in the following work. Finally, we have
achieved the accuracy of 99.94%, 98.63%, 96.59% in train-
ing, validation and testing set. Also, we obtained sensitiv-
ity and specificity of 99.93%, and 97.03% in the testing
set.
The experimental results have confirmed that the method
is effective for AF detection, especially in terms of accuracy,
sensitivity, and sensitivity. In addition, different from the
traditional feature learning methods used in the biomedical
signal processing, the proposed deep-CNN-BLSTM network
neither need the professional knowledge in the biomedi-
cal field nor the handcrafted feature extraction methods in
conventional machining learning. Otherwise, the network
accomplishes AF detection with little computation overhead,
which leads to real-time speed theoretically.
Our future research will involve the classification of multi-
ple arrhythmia signals and extend to other public databases
to assess the accuracy of the method and the generaliza-
tion ability of the model, as well as to optimize the learn-
ing ability of neural networks and enhance the clinical
contributions.
REFERENCES
[1] S. T. Mathew, J. Patel, and S. Joseph, ‘‘Atrial fibrillation: Mechanistic
insights and treatment options,’’ Eur. J. Internal Med., vol. 20, no. 7,
pp. 672–681, Nov. 2009.
[2] C. R. C. Wyndham, ‘‘Atrial fibrillation: The most common arrhythmia,’’
Texas Heart Inst. J., vol. 27, no. 3, pp. 257–267, Feb. 2000.
[3] V. Markides and R. J. Schilling, ‘‘Atrial fibrillation: Classification, patho-
physiology, mechanisms and drug treatment,’’ Heart, vol. 89, no. 8,
pp. 939–943, Aug. 2003.
[4] K. Harris, D. Edwards, and J. Mant, ‘‘How can we best detect atrial
fibrillation?’’ J. Roy. College Phys. Edinburgh, vol. 42, no. 18, pp. S5–S22,
Jan. 2012.
VOLUME 7, 2019
H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
[5] E. M. Hylek, A. S. Go, Y. Chang, N. G. Jensvold, L. E. Henault,
J. V. Selby, and D. E. Singer, ‘‘Effect of intensity of oral anticoagulation on
stroke severity and mortality in atrial fibrillation,’’ New England J. Med.,
vol. 349, no. 11, pp. 1019–1026, Sep. 2003.
[6] Y. Miyasaka, M. E. Barnes, B. J. Gersh, S. S. Cha, K. R. Bailey,
W. P. Abhayaratna, J. B. Seward, and T. S. M. Tsang, ‘‘Secular trends in
incidence of atrial fibrillation in Olmsted County, Minnesota, 1980 to 2000,
and implications on the projections for future prevalence,’’ Circulation,
vol. 114, no. 2, pp. 119–125, Jul. 2006.
[7] A. S. Gami, G. Pressman, S. M. Caples, R. Kanagala, J. J. Gard,
D. E. Davison, J. F. Malouf, N. M. Ammash, P. A. Friedman, and
V. K. Somers, ‘‘Association of atrial fibrillation and obstructive sleep
apnea,’’ Circulation, vol. 110, no. 4, pp. 364–367, Nov. 2004.
[8] M. Beed, ‘‘Bennett’s cardiac arrhythmias, practical notes on interpretation
and treatment,’’ Brit. J. Anaesthesia, vol. 112, no. 4, p. 775, Apr. 2014.
[9] K. Mukamal, J. Tolstrup, J. Friberg, G. Jensen, and M. Grønbæk, ‘‘Alcohol
consumption and risk of atrial fibrillation in men and women: The Copen-
hagen City heart study,’’ Circulation, vol. 112, no. 12, pp. 1736–1742,
Sep. 2005.
[10] X. Zhou, H. Ding, B. Ung, E. Pickwell-MacPherson, and Y. Zhang, ‘‘Auto-
matic online detection of atrial fibrillation based on symbolic dynamics
and Shannon entropy,’’ Biomed. Eng. Online, vol. 13, no. 1, pp. 13–18,
Dec. 2014.
[11] U. Desai, R. J. Martis, U. R. Acharys, C. G. Nayak, G. Seshikala, and
S. K. Ranjan, ‘‘Diagnosis of multiclass tachycardia beats using recurrence
quantification analysis and ensemble classifiers,’’ J. Mech. Med. Biol.,
vol. 16, no. 1, pp. 314–335, Feb. 2016.
[12] S. Asgari, A. Mehrnia, and M. Moussavi, ‘‘Automatic detection of
atrial fibrillation using stationary wavelet transform and support vector
machine,’’ Comput. Biol. Med., vol. 60, pp. 132–142, May 2015.
[13] J. R. Mehall, R. M. Kohut, Jr., E. W. Schneeberger, W. H. Merrill, and
R. K. Wolf, ‘‘Absence of correlation between symptoms and rhythm in
‘symptomatic’ atrial fibrillation,’’ Ann. Thoracic Surg., vol. 83, no. 6,
pp. 2118–2121, Jun. 2007.
[14] J. Mant, D. A. Fitzmaurice, F. D. Hobbs, S. Jowett, E. T. Murray, R. Holder,
M. Davies, and G. Y. Lip, ‘‘Accuracy of diagnosing atrial fibrillation on
electrocardiogram by primary care practitioners and interpretative diag-
nostic software: Analysis of data from screening for atrial fibrillation in
the elderly (SAFE) trial,’’ Brit. Med. J. (Clin. Res. Ed.), vol. 335, no. 7616,
pp. 380–382, Aug. 2007.
[15] M. D. Zeiler and R. Fergus, ‘‘Visualizing and understanding convolutional
networks,’’ presented at the ECCV, 2014.
[16] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
recognition,’’ presented at the CVPR, 2015.
[17] C. Liu, O. Julien, R. Erik, Q. Li, L. Zhao, N. Shamim, and C. D. Clifford,
‘‘A comparison of entropy approaches for AF discrimination,’’ Physiol.
Meas., vol. 39, no. 7, pp. 18–36, Jul. 2018.
[18] D. E. Lake and J. R. Moorman, ‘‘Accurate estimation of entropy in very
short physiological time series: The problem of atrial fibrillation detection
in implanted ventricular devices,’’ Amer. J. Physiol. Heart Circulat. Phys-
iol., vol. 300, no. 1, pp. H319–H325, Jan. 2011.
[19] N. Liu, M. Sun, L. Wang, W. Zhou, H. Dang, and X. Zhou, ‘‘A support
vector machine approach for AF classification from a short single-lead
ECG recording,’’ Physiol. Meas., vol. 39, no. 6, Jun. 2018, Art. no. 064004.
[20] L. Zhao, C. Liu, S. Wei, Q. Shen, F. Zhou, and J. Li, ‘‘A new entropy-
based atrial fibrillation detection method for scanning wearable ECG
recordings,’’ Entropy, vol. 20, no. 12, p. 904, Nov. 2018.
[21] Z. Mei, X. Gu, H. Chen, and W. Chen, ‘‘Automatic atrial fibrillation detec-
tion based on heart rate variability and spectral features,’’ IEEE Access,
vol. 6, pp. 53566–53575, Sep. 2018.
[22] A. Sološenko, A. Petrènas, B. Paliakaité, L. Sörnmo, and V. Marozas,
‘‘Detection of atrial fibrillation using a wrist-worn device,’’ Physiol. Meas.,
vol. 40, no. 2, Feb. 2019, Art. no. 025003.
[23] M. Shao, G. Bin, S. Wu, G. Bin, J. Huang, and Z. Zhou, ‘‘Detection of atrial
fibrillation from ECG recordings using decision tree ensemble with multi-
level features,’’ Physiol. Meas., vol. 39, no. 9, Sep. 2018, Art. no. 094008.
[24] V. Gliner and Y. Yaniv, ‘‘An SVM approach for identifying atrial fibrilla-
tion,’’ Physiol. Meas., vol. 39, no. 9, Sep. 2018, Art. no. 094007.
[25] N. Sadr, M. Jayawardhana, T. T. Pham, R. Tang, A. T. Balaei, and
P. De Chazal, ‘‘A low-complexity algorithm for detection of atrial fib-
rillation using an ECG,’’ Physiol. Meas., vol. 39, no. 9, Sep. 2018,
Art. no. 064003.
VOLUME 7, 2019
[26] M. Athif, P. C. Yasawardene, and C. Daluwatte, ‘‘Detecting atrial fibril-
lation from short single lead ECGs using statistical and morphological
features,’’ Physiol. Meas., vol. 39, no. 6, Jun. 2018, Art. no. 064002.
[27] T. Teijeiro, C. A. García, D. Castro, and P. Félix, ‘‘Abductive reasoning as a
basis to reproduce expert criteria in ECG atrial fibrillation identification,’’
Physiol. Meas., vol. 39, no. 8, Aug. 2018, Art. no. 084006.
[28] S. Parvaneh, J. Rubin, A. Rahman, B. Conroy, and S. Babaeizadeh, ‘‘Ana-
lyzing single-lead short ECG recordings using dense convolutional neural
networks and feature-based post-processing to detect atrial fibrillation,’’
Physiol. Meas., vol. 39, no. 8, Aug. 2018, Art. no. 084003.
[29] J. Harju, A. Tarniceriu, J. Tarniceriu, A. Vehkaoja, A. Yli-Hankala, and
I. Korhonen, ‘‘Monitoring of heart rate and inter-beat intervals with
wrist plethysmography in patients with atrial fibrillation,’’ Physiol. Meas.,
vol. 39, no. 6, Jun. 2018, Art. no. 065007.
[30] L. M. Eerikäinen, A. G. Bonomi, F. Schipper, L. R. C. Dekker,
R. Vullings, H. M. D. Morree, and R. M. Aarts, ‘‘Comparison between
electrocardiogram- and photoplethysmogram-derived features for atrial
fibrillation detection in free-living conditions,’’ Physiol. Meas., vol. 39,
no. 8, Aug. 2018, Art. no. 084001.
[31] I. Christov, V. Krasteva, I. Simova, T. Neycheva, and R. Schmid, ‘‘Ranking
of the most reliable beat morphology and heart rate variability features
for the detection of atrial fibrillation in short single-lead ECG,’’ Physiol.
Meas., vol. 39, no. 9, Sep. 2018, Art. no. 094005.
[32] M. Rizwan, B. M. Whitaker, and D. V. Anderson, ‘‘AF detection from ECG
recordings using feature selection, sparse coding, and ensemble learning,’’
Physiol. Meas., vol. 39, no. 12, Dec. 2018, Art. no. 124007.
[33] V. D. A. Corino, R. Laureanti, L. Ferranti, G. Scarpini, F. Lombardi, and
L. T. Mainardi, ‘‘Detection of atrial fibrillation episodes using a wristband
device,’’ Physiol. Meas., vol. 38, no. 5, p. 787, Apr. 2017.
[34] T. Conroy, J. H. Guzman, B. Hall, G. Tsouri, and J.-P. Couderc, ‘‘Detec-
tion of atrial fibrillation using an earlobe photoplethysmographic sensor,’’
Physiol. Meas., vol. 38, no. 10, p. 1906, Sep. 2017.
[35] R. J. Martis, U. R. Acharya, and H. Adeli, ‘‘Current methods in electro-
cardiogram characterization,’’ Comput. Biol. Med., vol. 48, pp. 133–149,
May 2014.
[36] U. R. Acharya, H. Fujita, S. L. Oh, Y. Hagiwara, J. H. Tan, and
M. Adam, ‘‘Application of deep convolutional neural network for auto-
mated detection of myocardial infarction using ECG signals,’’ Inf. Sci.,
vols. 415–416, pp. 190–198, Nov. 2017.
[37] M. M. Al Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. Melgani, and
R. R. Yager, ‘‘Deep learning approach for active classification of electro-
cardiogram signals,’’ Inf. Sci., vol. 345, pp. 340–354, Jun. 2016.
[38] U. R. Acharya, H. Fujita, S. L. Oh, Y. Hagiwara, J. H. Tan, and
M. Adam, ‘‘Automated detection of arrhythmias using different intervals of
tachycardia ECG segments with convolutional neural network,’’ Inf. Sci.,
vol. 405, no. 1, pp. 81–90, Sep. 2017.
[39] B. Pourbabaee, M. J. Roshtkhari, and K. Khorasani, ‘‘Deep convolutional
neural networks and learning ECG features for screening paroxysmal atrial
fibrillation patients,’’ IEEE Trans. Syst., Man, Cybern. Syst., vol. 48, no. 12,
pp. 2095–2104, Dec. 2018.
[40] X. Xu, S. Wei, C. Ma, K. Luo, L. Zhang, and C. Liu, ‘‘Atrial fibrillation
beat identification using the combination of modified frequency slice
wavelet transform and convolutional neural networks,’’ J. Healthcare Eng.,
vol. 2018, Jul. 2018, Art. no. 2102918.
[41] R. S. Andersen, A. Peimankar, and S. Puthusserypady, ‘‘A deep learning
approach for real-time detection of atrial fibrillation,’’ Expert Syst. Appl.,
vol. 115, pp. 465–473, Jan. 2019.
[42] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff,
P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. K. Peng, and
H. E. Stanley, ‘‘PhysioBank, PhysioToolkit, and PhysioNet: Components
of a new research resource for complex physiologic signals,’’ Circulation,
vol. 101, no. 23, pp. 215–220, Jun. 2000.
[43] C. Pater, ‘‘Methodological considerations in the design of trials for safety
assessment of new drugs and chemical entities,’’ Current Control Trials
Cardiovascular Med., vol. 6, no. 1, pp. 1–13, Feb. 2005.
[44] K. Fukushima, ‘‘Neocognitron: A self-organizing neural network model
for a mechanism of pattern recognition unaffected by shift in position,’’
Biol. Cybern., vol. 36, no. 4, pp. 193–202, Apr. 1980.
[45] J. J. Hopfield, ‘‘Neural networks and physical systems with emergent
collective computational abilities,’’ Proc. Nat. Acad. Sci. USA, vol. 79,
no. 8, pp. 2554–2558, Apr. 1982.
[46] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’ Neural
Comput., vol. 9, no. 8, pp. 1735–1780, Dec. 1997.
75589
 H. Dang et al.: Novel Deep Arrhythmia-Diagnosis Network for AF Classification Using ECG Signals
[47] F. A. Gers, J. Schmidhuber, and F. Cummins, ‘‘Learning to forget:
Continual prediction with LSTM,’’ Neural Comput., vol. 12, no. 10,
pp. 2451–2471, Oct. 2000.
[48] A. Radford, L. Metz, and S. Chintala, ‘‘Unsupervised representation learn-
ing with deep convolutional generative adversarial networks,’’ presented at
the ICLR, 2016.
[49] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro,
G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow,
A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur,
J. Levenberg, D. Mane, R. Monga, S. Moore, D. Murray, C. Olah,
M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker,
V. Vanhoucke, V. Vasudevan, F. Viegas, O. Vinyals, P. Warden, M. Watten-
berg, M. Wicke, Y. Yu, and X. Zheng, ‘‘TensorFlow: Large-scale machine
learning on heterogeneous systems,’’ 2016, arXiv:1603.04467. [Online].
Available: https://arxiv.org/abs/1603.04467
[50] A. Petrenas, V. Marozas, and L. Sörnmo, ‘‘Low-complexity detection
of atrial fibrillation in continuous long-term monitoring,’’ Comput. Biol.
Med., vol. 65, pp. 184–191, Oct. 2015.
[51] N. Henzel, J. Wróbel, and K. Horoba, ‘‘Atrial fibrillation episodes detec-
tion based on classification of heart rate derived features,’’ in Proc. 24th
Int. Conf. IEEE Mixed Design Integr. Circuits Syst. MIXDES, Jun. 2017,
pp. 571–576.
HAO DANG received the M.S. degree in pattern
recognition and intelligent system from the Henan
University of Technology, Zhengzhou, China,
in 2016. He is currently pursuing the Ph.D. degree
in control science and engineering with the School
of Automation, Beijing University of Posts and
Telecommunications, Beijing, China. His research
interests include the pattern recognition, intelli-
gent systems, and machine learning.
MUYI SUN received the B.S. degree in automa-
tion from the Beijing University of Posts and
Telecommunications, Beijing, China, in 2015,
where he is currently pursuing the Ph.D. degree in
control science and engineering with the School
of Automation. His research interests include the
pattern recognition and intelligent systems.
GUANHONG ZHANG received the B.S. degree
in automation from the Beijing University of Posts
and Telecommunications, Beijing, China, in 2015,
where he is currently pursuing the master’s degree
in pattern recognition and intelligent system from
the School of Automation. His research interests
include the machine learning, computer vision,
and neural networks.
75590
XINGQUN QI is currently pursuing the master’s
degree in control science and engineering from
the School of Automation, Beijing University of
Posts and Telecommunications, Beijing, China.
His research interests include the deep learning
and computer vision.
XIAOGUANG ZHOU received the M.S. degree
from the Department of Precision Instrument,
Tsinghua University, in 1984, and the Ph.D. degree
in engineering from the Tokyo University of Agri-
culture and Technology, Japan. He was a Visiting
Professor with the Tokyo University of Agriculture
and Technology, from 2001 to 2002, and a JSPS
Researcher with Tokyo University, from 2013 to
2014. He is currently a Professor and a Doc-
toral Supervisor with the School of Automation,
Beijing University of Posts and Telecommunications. He is also the Director
of the Engineering Research Center of Information Networks, Ministry
of Education. He is the author of over ten books, over 100 articles, and
over 16 inventions. His research interests include control theory and its
application in engineering, deep learning, computer vision, the Internet
of Things and automated logistics systems, and mechatronics technology.
He is also a permanent member of the Chinese Association of Automa-
tion/Manufacturing Technology Committee and the China Institute of Com-
munications/Equipment Manufacturing Technical Committee.
QING CHANG received the M.D. degree from the
Department of Clinical Medicine, Natong Med-
ical College, in 2000, and the Ph.D. degree in
molecular cell biology from The University of
Tokyo, in 2013. He was a Postdoctoral Fellow
with the Shanghai Jiao Tong University School of
Medicine, from 2013 to 2015. He is currently an
Associate Professor and a Master Supervisor with
the Shanghai University of Medicine & Health
Sciences. He is also an Attending Doctor and a
Research Fellow with the Shanghai General Practice Medical Education
and Research Center, Jiading District Central Hospital Affiliated Shanghai
University of Medicine & Health Sciences. He has published more than
20 articles. His research interests include the usage and challenge of inno-
vating technology in general practice medicine, such as sequencing, big data,
and AI.
VOLUME 7, 2019