Hindawi

Computational and Mathematical Methods in Medicine

Volume 2022, Article ID 2323625, 13 pages
https://doi.org/10.1155/2022/2323625

Research Article
Interpatient ECG Arrhythmia Detection by Residual
Attention CNN

Pengyao Xu ,1 Hui Liu,1 Xiaoyun Xie,1 Shuwang Zhou,1,2 Minglei Shu ,1

and Yinglong Wang 1
1
Shandong Artificial Intelligence Institute, Qilu University of Technology (Shandong Academy of Sciences), Jinan 250353, China
2
College of Computer Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, China

Correspondence should be addressed to Yinglong Wang; [email protected]

Received 4 January 2022; Revised 4 March 2022; Accepted 7 March 2022; Published 8 April 2022

Academic Editor: Zoran Bosnic

Copyright © 2022 Pengyao Xu et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The precise identification of arrhythmia is critical in electrocardiogram (ECG) research. Many automatic classification methods
have been suggested so far. However, efficient and accurate classification is still a challenge due to the limited feature
extraction and model generalization ability. We integrate attention mechanism and residual skip connection into the U-Net
(RA-UNET); besides, a skip connection between the RA-UNET and a residual block is executed as a residual attention
convolutional neural network (RA-CNN) for accurate classification. The model was evaluated using the MIT-BIH arrhythmia
database and achieved an accuracy of 98.5% and F 1 scores for the classes S and V of 82.8% and 91.7%, respectively, which is
far superior to other approaches.

1. Introduction ECG representation, i.e., one-dimensional (1-D) and two-

dimensional (2-D).
The latest survey statistics on global causes of mortality and Some studies exploit the original ECG signal as the
disability of the World Health Organization demonstrate model input. Although the proposed 1-D deep convolutional
that cardiovascular disease (CVD) is one of the most serious neural network (CNN) has achieved good classification
diseases that threaten human health. The ECG signal reflects results [8, 9], however, beat-by-beat classification cannot be
the electrical activity of the heart and is the primary basis for achieved due to the fixed time window size. Lin et al. [10]
the diagnosis of CVD. With the development of computer proposed a method based on normalized and nonnorma-
technology, automatic arrhythmia detection technology has lized RR intervals that extract ECG morphology by wavelet
become a research hotspot. analysis and linear prediction model, but this method
Traditional machine learning approaches such as inde- requires lots of signal preprocessing and has low prediction
pendent component analysis [1–3], principal component accuracy. Llamedo and Martínez [11] proposed a method
analysis (PCA) [4], support vector machine (SVM) [5], based on a linear classifier and a clustering algorithm; how-
and K-nearest neighbor (KNN) [6] have been utilized to ever, the clustering algorithm cannot effectively represent
identify arrhythmias. However, these methods require arti- class at the edge, making more likely arrhythmia misjudg-
ficial feature extraction and intervention. With the devel- ment. In addition, the abovementioned 1-D studies also
opment of technology, deep learning has gradually introduced a small degree of preprocessing.
become the mainstream method for automatic ECG classi- The ECG signal can also be converted from one dimen-
fication [7]. There are mainly two kinds of deep learning sion into two dimensions in various manners, such as fre-
approaches from the perspective of the dimension of quency spectrum and time-frequency images. Al Rahhal
2 Computational and Mathematical Methods in Medicine

et al. [12] use the continuous wavelet transform (CWT) to parallel, and then fuses the weight feature maps of
generate time-frequency information, then migration learn- the two branches together to guide model learning
ing. However, denoising and data augmentation operations
reduce model efficiency. Xia et al. [13] use the heartbeat The rest of this paper is organized as follows. The pro-
extraction method to convert multiple signals contained posed model is discussed in detail in Section 2, followed by
within 5 s into an image. However, the proposed structure the experimental design and verification in Section 3. Con-
not only limits the effect of the model due to the immutabil- clusions are finally drawn in Section 4.
ity of the short-time Fourier transform window but also eas-
ily causes misjudgment of normal data in verification 2. Methodology
because as long as one of the multiple heartbeats contained
2.1. Database. The suggested approach is trained and evalu-
in the image is abnormal, the entire image will be marked
ated using the MIT-BIH arrhythmia database [29]. It was
as abnormal. Li et al. [14] exploited three distinct types of
developed in collaboration between the Massachusetts Insti-
wavelet transforms paired with CNN to create a depth tech-
tute of Technology and Beth Israel Hospital in Boston and is
nique for automatically distinguishing time-frequency
now considered one of the three primary databases in aca-
images, which identified ventricular ectopic heartbeat (V)
demic circles. The database contains 48 Holter records from
as more than 97%; however, preprocessing operations such
25 men and 22 women between the ages of 32 and 89 (of
as noise reduction increase the complexity of the model.
which 201 and 202 are from the same male), all of which
Salem et al. [15] utilized DenseNet to classify ECG spectra
have significant variances. Each recording is a dual-channel
from the perspective of transfer learning, but it also has the
signal with a sampling rate of 360 Hz and a length of slightly
same risk of misjudgment as [13]. But in terms of overall
more than 30 minutes, with the R peak value of each heart-
performance, the 2-D ECG data is weaker than the 1-D sig-
beat indicated.
nal noise interference, which has also been proved in the
research [16, 17]. 2.2. Preprocessing
In order to solve the problems of cumbersome prepro-
cessing and difficult beat-by-beat classification in the above 2.2.1. ECG Heartbeat Segmentation. Because each heartbeat
research, inspired by structural variants such as fully convo- in an ECG has a distinct duration, the length of it segmented
lutional network, U-Net, residual network, and attention from an ECG is not equal. Different methods of heartbeat
mechanism [18–27] that have been successfully used in var- segmentation were employed in the literature [30–32] in
ious tasks (such as natural image classification and medical the study of 2-D.
image segmentation), this paper proposes an RA-CNN We directly used the R peak position in the MIT-BIH
model for the classification of arrhythmia between patients. database without additional positioning and confirmed the
Firstly, the CWT is used to convert the ECG heartbeat into beat length after positioning the QRS complex according to
an image and classes with much fewer samples are enhanced the R peak position [33]. Rcurrent , Rprevious , and Rlast represent
by data augmentation techniques. Secondly, the attention the R wave peaks of the currently located heartbeat and the
mechanism and residual skip connection are integrated into adjacent heartbeats before and after; the R-R interval
the U-Net which is called residual attention U-Net (RA- between two adjacent R waves is regarded as a segment. In
UNET). Finally, the RA-CNN constitutes by a skip connec- order to fully ensure the integrity of the segmented heartbeat
tion between the RA-UNET and a residual block. We medical information, the middle 3/4 position of the two R
trained and tested the models on the MIT-BIH database, peaks of Rprevious and Rlast is taken as the intercepted heart-
and the final experimental results demonstrate the superior- beat length; therefore, the intercepted n-th heartbeat can be
ity of the proposed method. expressed as Formula (1) (Figure 1):
The main advantages of the proposed method are sum-
marized as follows:

3 Rlast − Rprevious
EBeat = , ð1Þ
(1) The converted 2-D ECG will improve the effective 4
area that the model can learn and use data enhance-
ment methods to make up for the deficiency of where EBeat represent the extracted heartbeat, Rprevious and
waveforms [28]. The data enhancement on 1-D Rlast , respectively, represent the abscissa values of the previ-
ECG may change its time domain information, but ous and next heartbeat of the extracted heartbeat on the
this problem does not exist in 2-D images coordinate axis. If the extracted heartbeat has no heartbeat
(2) A new residual block (R-block) with judgment Rprevious or Rlast , the coordinates correspond to the heartbeat;
branches is proposed as the basic module of RA- then the current heartbeat will not be segmented.
CNN; it judges whether to retain the original feature
map and thus solves the performance degradation 2.2.2. Transforming the 1-D ECG into 2-D ECG. After deter-
mining the sampling length of each beat, the 1-D ECG is
(3) RA-UNET integrates the “split-transform-fusion” converted to the time-frequency domain by CWT [28].
principle, splits the feature map into two groups after The choice of CWT is motivated by its success at analyzing
each sampling operation, uses the two branches of ECG signals. The dimension of this output is higher than
spatial and channel generate attention weights in the dimension of the input. Unlike feature reduction,
Computational and Mathematical Methods in Medicine 3

Rprevious Rcurrent Rlast

Rcurrent Rprevious Rcurrent Rlast

4 4
R–R segment R–R segment

Figure 1: Heartbeat segmentation schematic diagram.

(i) (ii)
……

ECG original signal Generate2–D images heartbeat augmentation

Figure 2: 2-D data generation and enhancement.

Table 1: Data comparison before and after dataset enhancement.

Number of heart beats

Database Enhancement Type Total
N S V F
Amount 90042 2779 7007 802 100630
Before
Percentage (%) 89.48 2.76 6.96 0.80 —
MIT-BIH
Amount 90042 20696 41099 8668 160505
After
Percentage (%) 56.10 12.89 25.61 5.40 —

overcomplete representations allow finding more robust and after performing the preceding technique on the original
sparse feature representations from the data [12]. For ECG ECG. In Figure 2, step (i) depicts the process of turning
time series, its CWT relative to a given mother wavelet the extracted heartbeat into an image and step (ii) depicts
EBeat is defined as follows: a portion of the data augmentation impacts.
ð∞   The abovementioned heartbeat enhancement approach
1 t−b is utilized to improve the data in DS1 (introduced in detail
Ca,b EBeat ðt Þ = EBeat ðt Þψ dt: ð2Þ
j j1/2
a −∞ a in this work 3.1.3). Following processing, the data balance
is achieved in order to properly train the RA-CNN model.
Among them, a and b are the scale and translation Table 1 shows the number and percentage of heartbeats
parameters, respectively. EBeat ðtÞ is the given signal; ψ is before and after enhancement.
the mother wavelet.
2.3. Model Architecture. Figure 3 shows the overall flowchart
2.2.3. Heartbeat Augmentation. Even in patients with of the proposed RA-CNN model to classify arrhythmia. The
arrhythmia, the majority of the swings in the ECG analysis encoding as images module (left) is the preprocessing pro-
are normal signals, leading to fewer damage data in the cess in this work 2.2 to use CWT transform the 1-D ECG
ECG database. The use of data augmentation techniques to into 2-D ECG heartbeat. The RA-CNN model (middle) is
boost damage data can effectively make up for the absence designed to learn 2-D ECG features so as to transform it to
of training data. Decrease the danger of overfitting, and the forms that easy to classify. The arrhythmia prediction
increase the algorithm’s robustness. module (right) realizes the classification in terms of the out-
According to the characteristics of the 2-D ECG wave- put of RA-CNN according to arrhythmias in the AAMI
form, this article will move the beat to the left and right, standard.
move up, and move down to obtain multiple enhanced The RA-CNN model consists of three parts: top layer,
heartbeat images. The signal characteristics in the original middle layer, and bottom layer (as shown in Figure 4). The
ECG can be significantly retained by using the augmented left part of the top layer uses conv2d, avg pooling, and R-
images [34–36]. Multiple focal heartbeat data can be created block to perform a certain degree of feature reduction on
4 Computational and Mathematical Methods in Medicine

Encoding as Images RA–CNN Model Arrhythmia Predication

Input

⊕
N

Pre–processing F

ECG heartbeat segmentation S

Generate 2-D Images ⊗
Heartbeat Augmentation Output

R–block Max pooling

⊕ ⊕
A–block Conv2d

Upsampling Avg pooling

⊗ Matrix multiplication
⊕ Element-wise sum

Figure 3: RA-CNN model training flowchart.

model can efficiently focus on the appropriate area of the

ECG feature. At the same time, each output of downsam-
Top layer (left) Top layer (right)
pling is used as a carrier to save the characteristics of the fea-
ture map via the skip connection with the output of the
⊕ upsampling in the same size, which prevents inaccurate fea-
ture reconstruction. The number of image channels and size
changes in the RA-CNN model structure are shown in
Table 2.
⊗ Middle
layer (1) Residual block (R-block): it is an encapsulated resid-
ual module with several convolution layers as the
⊕ ⊕
network infrastructure; it performs general feature
Bottom learning operations or dimensionality reduction
layer operations (such as 2.3.1).
(2) Residual Attention UNET (RA-UNET): it includes a
complete downsampling and upsampling process
Figure 4: RA-CNN model. through the hourglass structure; the module has fully
learned the inherent characteristics of 2-D ECG. RA-
UNET converts the intrinsic feature map output of
the 2-D ECG image, which is conducive to reducing the size each upsampling into an attention mask to guide
of the input (record the output as initial feature map) and the feature learning of the model through skip con-
expanding the receptive fields. The right part of the top layer nection, so that the model can suppress the worthless
reduces the image dimension to 1 × 1 in order to classify by area of the feature map while enhancing specific
multiple consecutive R-block and avg pooling. The skip con- important information (such as 2.3.2).
nection in the top layer is to connect the initial feature map (3) Attention block (A-block): channel attention and
and the output features of the other two layers. In the middle spatial attention are learned in parallel by grouping
layer, the initial feature map passes through only an R-block feature maps along the channel axis to achieve more
and then connects with the output of the bottom layer. The accurate attention to important information areas
bottom layer is residual attention U-Net (RA-UNET) which (such as 2.3.3).
is an hourglass structure from top-to-bottom to bottom-to-
top, i.e., from downsampling to upsampling; the downsam-
pling is achieved by R-block that extracts the essential fea- 2.3.1. R-Block. R-block is a basic residual block with judg-
tures from high-dimensional images and upsampling to be ment branches, which is made up of three BatchNorm2d-
done by bilinear interpolation. A-block is applied after each Relu-Conv2d layers and then distributed throughout the
downsampling and upsampling to intensify the output by RA-CNN model to accomplish the general function of fea-
generating the attention weight distribution, so that the ture processing.
Computational and Mathematical Methods in Medicine 5

Table 2: The number of channels and output dimensions of each layer.

Layer name Operate Kernel size Stride Output size Channels

Input 224 × 224 3
conv2d 7×7 2 112 × 112 16
Max Pool2d 3×3 2 56 × 56 16
0 1
Top layer B C 1
B conv2d, 1 × 1, 4 C
B C
R-block1 B conv2d, 3 × 3, 4 C × 1 1 56 × 56 16
@ A 1
conv2d, 1 × 1, 16
0 1
B C 1
B conv2d, 1 × 1, 4 C
B C
Middle layer R-block5 B conv2d, 3 × 3, 4 C × 1 1 56 × 56 16
@ A 1
conv2d, 1 × 1, 16
Bottom layer RA-UNET 56 × 56 16
0 1
B C 1
B conv2d, 1 × 1, 8 C
B C
R-block2 B conv2d, 3 × 3, 8 C × 1 2 28 × 28 32
@ A 1
conv2d, 1 × 1, 32
0 1
B C 1
B conv2d, 1 × 1, 16 C
B C
R-block3 B conv2d, 3 × 3, 16 C × 1 2 14 × 14 64
Top layer @ A 1
conv2d, 1 × 1, 64
0 1
B C 1
B conv2d, 1 × 1, 16 C
B C
R-block4 B conv2d, 3 × 3, 16 C × 1 2 7×7 64
@ A 1
conv2d, 1 × 1, 64
Avg Pool2d 7×7 1 1×1 64
Output 4

Figure 5 shows the structural details of the R-block,

Input
which was inspired by the ResNet to solve the “degradation”
problem caused by very deep levels and designed a structure
BN with a judgment function (the Exit? branch shown in
Relu Figure 5), which decides whether to retain more original fea-
Conv2d ture information by setting different steps and channels, so
the purpose of it is to ensure that the essential characteristics
BN of the feature map will not be destroyed to the maximum
Relu
extent. Therefore, we can set appropriate parameters for dif-
ferent needs, followed by the residual connection.
Conv2d (ii) (i)
For the input X R of R-block, the expected output RðX R Þ
can be expressed as
BN
Relu 8  
>
< f i θ ðσ1 ðX R ÞÞ ⊕ X R ðiÞ,
T
Conv2d
R ðX R Þ =    
>
: f θT ðσ ðX ÞÞ ⊕ f θT ðσ ðX ÞÞ ðiiÞ,
⊕ Exit? i 1 R 1 1 R

ð3Þ
Output

Figure 5: R-block. where f i ð∙Þ is the ith BatchNorm2d-Relu-Conv2d opera-

tion, θ is a convolution operation, σ1 ð∙Þ is a ReLU function,
6 Computational and Mathematical Methods in Medicine

and ⊕ denotes the element-wise sum. In the Exit? process of

judgment, when the number of input and output channels is
equal or the convolution step is 1, the flow is shown in pro- ⊕ ⊕
cess ðiÞ of Figure 5 and the expected output RðX R Þ is shown
in the formula 3-ðiÞ. If not, the flow is shown in process ðiiÞ
in Figure 5 and the expected output RðX R Þ is shown in for-
mula 3-ðiiÞ. Then the final output feature map RðX R Þ ∈
RC×H×W .
The R-block solves the problem of degradation and gra-
dient disappearance through the residual connection with
judging branches, which improves the network performance
and reduces the feature dimension by changing the number
R–block Upsampling
of channels or stride in the branch structure.
A–block ⊕ Element–wise sum
2.3.2. RA-UNET. RA-UNET is an improvement of the U-
Max pooling
Net [18–22] by incorporating residual and attention mecha-
nisms. RA-UNET is an encoder-decoder structure (as shown Figure 6: RA-UNET structure diagram.
in Figure 6), which extracts high-level information based on
three layers of downsampling and then reconstructs the fea-
ture by three layers of upsampling. In our design, the most
significant thing is the attention block (A-block) inserted channel axis, and each group of features is split into two
after each downsampling and upsampling, which can assist branches for channel attention and spatial attention, respec-
the model in accurate and efficient feature reduction and tively, and then concatenates the attention results of the two
reconstruction. We will introduce its implementation in branches together. Finally, the n groups of features are
detail: merged to obtain a feature map with the same size as the
input. Figure 7 shows in detail one group of attention mech-
(i) Encoder: using max pooling to realize the resample anisms after channel grouping.
of vital information of the input image, i.e., down Take the feature map X ∈ RC×H×W as an example, which
sampling, at the same time, the A-block is used to is the output after the first use of max pooling in RA-UNET.
strengthen the effect of key areas. First, divide its channel dimension into n groups of subfea-
tures X i ∈ Rðc/nÞ×H×W ð1 ≤ i ≤ n); then split each subfeature
(ii) Decoder: the upsampling operation is accomplished
through the bilinear interpolation layer, which can along the channel axis into two branches X i1 , X i2 ∈
be intuitively understood as the restoration process Rðc/2nÞ×H×W ð1 ≤ i ≤ nÞ; hence, the channel attention is per-
of the feature map. After each step of the upsam- formed on the first branch to embed global information
pling operation, the A-block is also used to encour- and generate channel statistical attention weight distribution
age the model to use the learned knowledge to learn by average pooling layer and softmax function. Then, the
more feature map information. channel attention weight distribution is imposed on X i1 to
help model focus on the distinct channel, followed with the
(iii) Skip connections: in order to better train the deep residual connection. The final output feature map X ′i1 of
network, after downsampling and completing the the channel attention can be realized as follows:
A-block, the R-block for feature processing not only
better integrates contextual semantic features and
prevents the disappearance of gradients caused by
the stacking of coding layers but also acts as a car- X ′i1 = ðσ2 ðW 1 ∙AVGðX i1 Þ + b1 Þ ⊗ X i1 Þ ⊕ X i1 : ð4Þ
rier to save the characteristics; it can better restore
the details of the same size feature map during the
upsampling process, so as to improve the recogni-
tion effect of the network on the diversity of wave- Among them, σ2 ð∙Þ represents the softmax function,
form changes.
AVG (∙) is the average pooling operation, W 1 ∈ Rðc/2nÞ×1×1
The specific size changes and convolution kernel size and b1 ∈ Rðc/2nÞ×1×1 are parameters used for scaling and
during RA-UNET processing are shown in Table 3. translation, and ⊗ stands for matrix multiplication.
Next, the spatial attention is performed on the second
2.3.3. A-Block. A-block captures remote contextual informa- branch to generate the spatial attention map which pays
tion in the spatial dimension and channel dimension, more attention to the important pixel area that stands
respectively. The attention mechanism is an improvement for the principal character of the feature map. What is dif-
in the article [24], which is used to automatically learn and ferent from channel attention is that X i2 obtained the spa-
calculate the contribution of input data to output data. First, tial attention weight distribution via group normalization,
the sampled feature map is divided into n groups along the and other operations are similar. The final output feature
Computational and Mathematical Methods in Medicine 7

Table 3: Each layer structure and input size of RA-UNET.

Name Layer Kernel size Output size Channels

Max Pool2d 3 × 3, stride 2 28 × 28 16
A-block — 28 × 28 16
0 1
B C
B conv2d, 1 × 1, 4 C
B C
R-block B conv2d, 3 × 3, 4 C × 1 28 × 28 16
@ A
conv2d, 1 × 1, 16
Max Pool2d 3 × 3, stride 2 14 × 14 16
Encoder
A-block — 14 × 14 16
0 1
B C
B conv2d, 1 × 1, 4 C
B C
R-block B conv2d, 3 × 3, 4 C × 1 14 × 14 16
@ A
conv2d, 1 × 1, 16
Max Pool2d 3 × 3, stride 2 7×7 16
A-block — 7×7 16

Upsample Size (14, 14) 14 × 14 16

A-block — 14 × 14 16
Upsample Size (28, 28) 28 × 28 16
A-block — 28 × 28 16
Upsample Size (56, 56) 56 × 56 16
Decoder
A-block — 56 × 56 16
0 1
B C
B conv2d, 1 × 1, 4 C
B C
R-block B conv2d, 3 × 3, 4 C × 1 56 × 56 16
@ A
conv2d, 1 × 1, 16

Output 56 × 56 16

X′i1

Xi1 ⊗ ⊕ W
c
Channel 2n h
c
h Xi 2n X′i
c c
W n n
⊗ ⊕
Xi2
c c
2n X′i2 2n
Spatial

Figure 7: A-block.

map X ′i2 of spatial attention can be achieved as follows: Among them, GNð∙Þ denotes the group normalization,
W 2 ∈ Rðc/2nÞ×1×1 and b2 ∈ Rðc/2nÞ×1×1 are model parameters
need to be trained.
X ′i2 = ðσ2 ðW 2 ∙GN ðX i2 Þ + b2 Þ ⊗ X i2 Þ ⊕ X i2 : ð5Þ In order to maintain the consistency of channel dimen-
sions after the attention operation, the channel attention
8 Computational and Mathematical Methods in Medicine

Table 4: Classification of ECG in the MIT-BIH database using AAMI standard.

Types Contains heartbeat type

Normal (NOR), left bundle branch block (LBBB), right bundle branch block (RBBB), atrial
Normal (N)
escape (AE), node (junction) escape heartbeat (NE)
Ventricular ectopic heartbeat (V) Premature ventricular contraction (PVC), ventricular escape heartbeat (VE)
Fusion heartbeat (F) Fusion of ventricular and normal (FVN)
Supraventricular ectopic heartbeat or Atrial premature (AP), aberrant atrial premature (AaP), nodal (junctional) premature (NP),
premature heartbeat (S) supraventricular premature (SP)
Unknown heartbeat (Q) Paced (/), fusion of paced and normal (FPN), unclassified (U), undetermined (?)

Table 5: Interpatient dataset partitioning scheme.

Number of heart beats

Database Datasets Partition Total
N S V F
Training 45824 18860 37880 8280 110844
DS1
MIT-BIH Percentage (%) 41.34 17.01 34.17 7.47 100
DS2 Testing 44218 1836 3219 388 49661
Total 90042 20696 41099 8668 160505

feature map and the spatial attention feature map are spliced classification standard, ignoring 4 records with severe noise
along the channel axis. among the 48 records. For the remaining records, an inter-
n o patient division scheme proposed in [37–42] is used. Divide
X i′ = Concat X ′i1 , X ′i2 , ð6Þ into training set (DS1) and test set (DS2). DS1 contains 22
records for training and parameter determination. DS2 is
only used as a test set for final performance evaluation.
where Concat {∙} denotes the dimension concatenating Using this partitioning method, there is no need to worry
operation and X i′ ∈ Rðc/nÞ×H×W ð1 ≤ i ≤ nÞ. about including the same patient’s heartbeat in both training
Finally, after n groups of feature maps are also aggre- and test sets. The number of heart beats after division is
gated along the channel dimension, the final attention fea- shown in Table 5.
ture map containing the weight coefficient is generated:
X ′ = Concat fX 1′ , X 2′ , ⋯X n′ g.
3.1.4. Training Parameter Setting. The learning rate is a key
2.3.4. Arrhythmia Predication. Finally, the RA-CNN model training parameter in the proposed RA-CNN model. We
uses a fully connected layer to perform a fully connected optimize the parameters in order to train the model for the
operation on the learned attention feature map to achieve best performance in arrhythmia classification.
arrhythmia classification. We set the initial learning rate to 0.001 and drop to the
original 0.1 every 20 epochs. In order to reduce the memory,
3. Experimental Design use a smaller batch size for training, and set the batch size to
a small batch of 16; the loss function uses cross entropy
3.1. Experimental Setup error, and the optimization function uses Adam.
3.1.1. Experimental Environment. The data preparation sec-
tion of this paper is done on an i7-10700K processor. The 3.1.5. Evaluation Metrics. This study utilized the MIT-BIH
experiment was done with the NVIDIAA 100 graphics card arrhythmia database to evaluate the RA-CNN model accord-
and completed on the Ubuntu 18.04.3 operating system. Run ing to the AAMI standard in order to test its performance.
PyTorch, and then use WFDB packet to process the ECG These indicators have also been employed extensively in
signal. research [37–42]: classification accuracy (Acc), sensitivity
(Sen), positive prediction rate (Ppr), and F 1 -score.
3.1.2. Classification Standard of ECG. This study used the Acc is the proportion of correctly classified ECG samples
widely used [37–42] American progressive association to the total sample and is also the most commonly used eval-
AAMI to develop medical device ANSI/AAMI EC57:2012 uation index in all classification problems.
standards to classify arrhythmias. Arrhythmias are divided
into five classes, as shown in Table 4.
3.1.3. Database Set. The data from MIT-BIH is used to train TP + TN
Acc = × 100%: ð7Þ
the model in this work. This paper strictly follows the AAMI TP + TN + FP + FN
Computational and Mathematical Methods in Medicine 9

(i) Original image

(ii) Channel attention image

(iii) Spatial attention image

Figure 8: 2-D ECG data processing results of the two branches of A-block.

Sen only processes positive heartbeats, which means the

40000
ratio of the detected true positive heartbeats to the actual N 43406 435 377 0
positive heartbeats. 35000

30000
TP
Sen = × 100%: ð8Þ S 317 1394 125 0
True label

TP + FN 25000

20000
Ppr represents the proportion of positive heartbeats that
V 240 84 2895 0
are correctly detected among all positive heartbeats. 15000

10000
TP
Ppr = × 100%: ð9Þ F 197 35 93 36 5000
TP + FP
0
F 1 -score is a comprehensive evaluation index of preci- N S V F
sion rate and recall rate, used to reflect the overall situation.
Predicted label

2 × Sen × Ppr Figure 9: Confusion matrix without data augmentation.

F1 = × 100%: ð10Þ
Sen + Ppr
40000
Among the above four evaluation indicators, false posi- N 43718 284 215 1
tive (FP) is the number of heartbeats that are misclassified. 35000
For example, it is actually a heartbeat of class N but is clas- 30000
sified into one of the classes V, F, or S. False negative (FN) is S 242 1525 69 0
True label

the number of heartbeats classified in different categories; it 25000

is also a misclassification of samples. True positive (TP) is 20000
the number of heartbeats that are correctly classified. True
V 174 36 3002 0 15000
negative (TN) is the number of heartbeats that do not belong
to a certain category and are not classified as such. 10000
3.2. Experimental Verification F 189 4 48 147 5000

3.2.1. Analysis of the Impact of A-Block on Classification 0

Results. Figure 8 shows the heartbeat display of channel N S V F
attention and spatial attention after A-block processing in Predicted label
the process of using RA-UNET. A-block explores attention
by assigning higher weights to pixels that are helpful for Figure 10: Confusion matrix enhanced with data augmentation.
accurate classification. Therefore, as the depth of the RA-
UNET deepens, the pixel area that represents the ECG curve
in the feature map will become more and more obvious. The
10 Computational and Mathematical Methods in Medicine

Table 6: Comparison of effects before and after data enhancement.

N (%) S (%) V (%)

Enhancement ACC
SEN Ppr F1 SEN Ppr F1 SEN Ppr F1
Without 97.6 98.16 98.29 98.23 75.93 71.56 73.68 89.93 82.95 86.30
Proposed 98.5 98.87 98.64 98.75 83.06 82.48 82.77 93.46 90.04 91.72

Table 7: Data analysis of ablation experiments.

N (%) S (%) V (%)

Works ACC
SEN Ppr F1 SEN Ppr F1 SEN Ppr F1
Without R-block 97.4 97.49 98.41 97.94 77.72 71.85 74.67 91.92 78.26 84.54
Without A-block 97.2 97.72 98.15 97.94 71.35 66.23 68.69 88.38 78.77 83.30
Without channel attention 97.7 98.18 98.21 98.20 76.68 68.84 72.61 89.84 86.35 88.06
Without spatial attention 97.5 97.95 98.08 98.02 75.44 68.40 71.74 88.75 83.51 86.05
Without top layer 96.5 96.36 97.88 97.11 74.83 64.87 69.50 90.59 72.86 80.76
Without middle layer 97.3 97.28 98.48 97.88 80.39 72.60 76.30 92.17 77.19 84.02
Proposed 98.5 98.87 98.64 98.75 83.06 82.48 82.77 93.46 90.04 91.72

RA-UNET model will not only focus more precisely on the overlap more in the distribution, and the classification effect
specific area of the lower part of the image where the wave- is not ideal.
form changes more but also filter the background informa-
tion. Thereby, it can “do no useless work” and has the 3.2.3. Ablation Study. It has been proved by 3.2.2 that the
effect of improving the classification accuracy. In the figure, data enhancement method proposed in this work is effective.
(i) shows 8 beats randomly selected from 2-D ECG, (ii) Therefore, the effectiveness of the proposed two basic struc-
shows the visualization results output by Channel attention tures of R-block and A-block is verified in the same situation
in A-block for the first time, and (iii) shows the output result using the enhancement method proposed in this work.
of spatial attention structure processing. Obviously, it can be Table 7 presents the results of our ablation experiments.
seen that (iii) pays more attention to the lower area of the First of all, we verify the influence of the R-block module
image than (ii) and realizes that the large-scale, multichan- on the model effect. We use conv2d (the same as the conv2d
nel features are concentrated in the key positions of the var- used in R-block) to replace the R-block that implements the
ious waveforms at the bottom of the image. downsampling effect in the model and remove the R-block
that implements the general feature processing function.
3.2.2. Data Enhancement Experiment. Figures 9 and 10, The final implementation result (as shown without R-block)
respectively, show the best results of classification of classes shows that the classification effect would be reduced without
N, S, V, and F ECG using RA-CNN when only setting var- R-block, so R-block is effective for improving the classifica-
iables for data enhancement. It can be found that the num- tion effect.
ber of correctly classified samples after enhancement has Secondly, we verify the effectiveness of A-block. First,
increased compared with that before enhancement. remove the A-block used to capture contextual information
Table 6 shows the evaluation results before and after data after the sampling step. The experimental results show that
enhancement using the indicators mentioned in 3.1.5. It can A-block also has a greater impact on the accuracy of classi-
be seen that with the basic settings unchanged, the average fication. Then, the effectiveness of the channel attention
accuracy of the data enhancement method proposed in this branch and the spatial attention branch in the A-block were
work has increased by about 0.8%. Other indicators have verified. By removing the two branches, respectively, it was
also improved, so the data enhancement method proposed proved that the two branches also have an important influ-
in this work can promote the classification results. ence on the context information capture of the A-block,
The final experimental results show that the model has a through the evaluation of the three classes of N, S, and V
good classification effect on class N and class V, while the through the general evaluation indicators.
class S classification effect is significantly lower than the Finally, we verify the effectiveness of the skip connection
other two classes. The main reason is that the number of used in the top layer and middle layer. The reason why the
training samples for class S is significantly less than the other skip connection structure is used is that RA-UNET uses
two categories even with data enhancement. The second is the function of ReLU in the feature learning process, which
that the similarity of the waveforms between class S and class will make the output result between (0, 1); therefore, the
N is extremely high, causing the two types of samples to value of the feature map will decrease over time as a result
Computational and Mathematical Methods in Medicine 11

Table 8: Comparison of related experiments.

N (%) S (%) V (%)

Works ACC
SEN Ppr F1 SEN Ppr F1 SEN Ppr F1
Dictionary
95.1 90.9 99.4 94.2 80.8 48.8 60.8 82.2 85.4 83.8
(2018) [38]
DCNN
94.0 90.6 98.8 94.5 82.3 38.1 52.1 92.0 72.1 80.9
(2018) [39]
MPCNN
96.4 98.8 97.4 98.1 76.5 76.6 76.6 85.7 94.1 89.7
(2019)[40]
DDCNN + CLSM (2020) [41] 95.1 97.5 97.6 97.6 83.8 59.4 69.5 80.4 90.2 85.0
Linear discriminant (2021) [42] 87.3 78.7 99.3 87.8 89.4 37.5 52.9 86.5 93.0 89.6
Proposed 98.5 98.9 98.6 98.8 83.1 82.5 82.8 93.5 90.1 91.7

of a series of feature learning operations, resulting in unsat- trocardiograms, they are mostly obtained in the form of
isfactory learning effects. Through the addition of the rela- images, and two-dimensional research is more conducive
tively original features of the top layer and middle layer, it to visualization, thereby improving the efficiency of diagno-
is possible to minimize the loss of important information sis and prevention of CVD. The data does not require any
without attention learning. The final experimental findings form of noise reduction operation and manual feature
also fully validate the efficacy of this step. extraction, which avoids the loss of detailed information in
the original ECG data and affects the feature extraction effect
3.2.4. Performance Comparison. We compared this study to [16, 17]. The preprocessing does not need to strictly extract a
similar studies in recent years to verify the advanced nature single heartbeat. Even if the heartbeat is mixed with the
of RA-CNN in the classification of arrhythmia. Table 8 dis- information of the front and back heartbeats, the ECG char-
plays the research findings based on data from the MIT-BIH acterization information can be better expressed through the
arrhythmia database, which has been segmented in the same CWT, and finally, a good classification performance can be
way as this paper. Each method’s name, the year it was pro- achieved.
posed, and its performance in the classification task are In a further work, we will investigate the improved ECG
listed in the table. network and further improve the classification performance
[38] used traditional methods for classification research, of different types of diseases [43–46]. On the clinical side, we
introduced 60 features for the classification step. Not only will develop an ECG system that can be deployed on wear-
was the preprocessing process complicated, but also the class able medical devices and automatic diagnosis algorithm, test,
S Ppr value was 48.8%, which is not ideal. [39] It is necessary and improve its performance [9, 47].
to read multiple heartbeat features for heartbeat classifica-
tion, which undoubtedly increases the amount of calcula- Data Availability
tion. [40] In addition to inputting the original signal as
input, the model also introduces RR interval information, The ECG signal data used to support the findings of this
which requires additional feature extraction operations, study have been deposited in the MIT-BIH Arrhythmia
and the obtained classification effect is also worse than this Database repository (https://www.physionet.org/content/
study [41]. After completing the initial classification using mitdb/1.0.0/).
a deep dual-channel CNN (DDCNN), it is necessary to fur-
ther use the central-towards LSTM supportive model Conflicts of Interest
(CLSM) to distinguish classes N and S; however, the classifi-
cation effect of category S is still unsatisfactory. [42] not only There are no conflicts of interest declared by the authors.
performed tedious noise reduction processing but also intro-
duced the RR interval relationship as a feature for learning, Acknowledgments
which undoubtedly increased the difficulty of feature extrac- This work was supported by the Major Research Plan of
tion. Compared with the above experiments, this model not Shandong Province: Research and Integrated Application
only has a simple feature extraction process but also has a of Key Technologies for Smart Healthcare (grant number
higher F 1 value for beat-by-beat classification, which is supe- 2020CXGC010903).
rior in class S pathology identification [38–42].
References
4. Conclusion
[1] J. P. Sahoo, Analysis of ECG signal for detection of cardiac
In this work, we propose a novel and effective RA-CNN arrhythmias, National Institute Of Technology, Rourkela,
model. Experiments on arrhythmia data interpatients show 2011.
that the model has a high ECG recognition ability, strong [2] T. F. Romdhane and M. A. Pr, “Electrocardiogram heartbeat
generalization, and robustness. When doctors diagnose elec- classification based on a deep convolutional neural network
12 Computational and Mathematical Methods in Medicine

and focal loss,” Computers in Biology and Medicine, vol. 123, [18] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional net-
p. 103866, 2020. works for semantic segmentation,” in Proceedings of the IEEE
[3] X. Jiang, L. Zhang, Q. Zhao, and S. Albayrak, “ECG arrhyth- conference on computer vision and pattern recognition, 2015.
mias recognition system based on independent component [19] O. Ronneberger, P. Fischer, and T. Brox, U-net: convolutional
analysis feature extraction,” in TENCON 2006-2006 IEEE networks for biomedical image segmentation, Springer, Cham,
Region 10 Conference, 2006. 2015.
[4] R. J. Martis, U. R. Acharya, C. M. Lim, and J. S. Suri, “Charac- [20] H. Huang, L. Lin, R. Tong et al., “Unet 3+: a full-scale con-
terization of ECG beats from cardiac arrhythmia using discrete nected UNet for medical image segmentationC//ICASSP,” in
cosine transform in PCA framework,” Knowledge-Based Sys- ICASSP 2020-2020 IEEE International Conference on Acous-
tems, vol. 45, pp. 76–82, 2013. tics, Speech and Signal Processing, pp. 1055–1059, 2020.
[5] M. Suchetha and N. Kumaravel, “Classification of arrhythmia [21] S. Guan, A. A. Khan, S. Sikdar, and P. V. Chitnis, “Fully dense
in electrocardiogram using EMD based features and support UNet for 2-D sparse photoacoustic tomography artifact
vector machine with margin sampling,” International Journal removal,” IEEE Journal of Biomedical and Health Informatics,
of Computational Intelligence and Applications, vol. 12, no. 3, vol. 24, no. 2, pp. 568–576, 2020.
p. 1350015, 2013. [22] S. Wu, Z. Wang, C. Liu, C. Zhu, S. Wu, and K. Xiao, “Automa-
[6] J. Park, K. Lee, and K. Kang, “Arrhythmia detection from tical segmentation of pelvic organs after hysterectomy by using
heartbeat using k-nearest neighbor classifier,” in 2013 IEEE dilated convolution u-net++,” in International Conference on
International Conference on Bioinformatics and Biomedicine, Software Quality, Reliability and Security Companion,
pp. 15–22, Dec. 2013. pp. 362–367, 2019.
[7] V. Gupta, N. K. Saxena, A. Kanungo, A. Gupta, P. Kumar, and [23] O. Russakovsky, J. Deng, H. Su et al., “Imagenet large scale
Salim, “A review of different ECG classification/detection tech- visual recognition challenge,” International Journal of Com-
niques for improved medical applications,” International Jour- puter Vision, vol. 115, no. 3, pp. 211–252, 2015.
nal of System Assurance Engineering and Management, pp. 1– [24] Q. L. Zhang and Y. B. Yang, “Sa-net: shuffle attention for deep
15, 2022. convolutional neural networks,” in ICASSP 2021-2021 IEEE
[8] S. Kiranyaz, T. Ince, and M. Gabbouj, “Real-time patient- International Conference on Acoustics, Speech and Signal Pro-
specific ECG classification by 1-D convolutional neural net- cessing, pp. 2235–2239, 2021.
works,” IEEE Transactions on Biomedical Engineering, [25] N. Soans, E. Asali, Y. Hong, and P. Doshi, “Sa-net: robust state-
vol. 63, no. 3, pp. 664–675, 2016. action recognition for learning from observations,” in Interna-
[9] H. D. M. Ribeiro, A. Arnold, J. P. Howard et al., “ECG-based tional Conference on Robotics and Automation, pp. 2153–2159,
real-time arrhythmia monitoring using quantized deep neural 2020.
networks: a feasibility study,” Computers in Biology and Med- [26] F. Wang, M. Jiang, C. Qian et al., “Residual attention network
icine, vol. 143, p. 105249, 2022. for image classification,” in Proceedings of the IEEE conference
[10] C. C. Lin and C. M. Yang, “Heartbeat classification using nor- on computer vision and pattern recognition, pp. 3156–3164,
malized RR intervals and morphological features,” Mathemat- 2017.
ical Problems in Engineering, vol. 2014, 11 pages, 2014. [27] S. M. A. Sharif, R. A. Naqvi, and M. Biswas, “Learning medical
[11] M. Llamedo and J. P. Martínez, “An automatic patient- image denoising with deep dynamic residual attention net-
adapted ECG heartbeat classifier allowing expert assistance,” work,” Mathematics, vol. 8, no. 12, p. 2192, 2020.
IEEE Transactions on Biomedical Engineering, vol. 59, no. 8, [28] R. He, K. Wang, N. Zhao et al., “Automatic detection of atrial
pp. 2312–2320, 2012. fibrillation based on continuous wavelet transform and 2D
[12] M. M. Al Rahhal, Y. Bazi, M. Al Zuair, E. Othman, and convolutional neural networks,” Frontiers in Physiology,
B. BenJdira, “Convolutional neural networks for electrocardio- vol. 9, p. 1206, 2018.
gram classification,” Journal of Medical and Biological Engi- [29] R. Mark and G. Moody, MIT-BIH Arrhythmia Database Direc-
neering, vol. 38, no. 6, pp. 1014–1025, 2018. tory, Cambridge, MA, MIT, USA, 1988.
[13] Y. Xia, N. Wulan, K. Wang, and H. Zhang, “Detecting atrial [30] K. Luo, J. Li, Z. Wang, and A. Cuschieri, “Patient-specific deep
fibrillation by deep convolutional neural networks,” Com- architectural model for ECG classification,” Journal of
puters in Biology and Medicine, vol. 93, pp. 84–92, 2018. Healthcare Engineering, vol. 2017, Article ID 4108720, 13
[14] Q. Li, C. Liu, Q. Li et al., “Ventricular ectopic beat detection pages, 2017.
using a wavelet transform and a convolutional neural net- [31] G. Yan, S. Liang, Y. Zhang, and F. Liu, “Fusing transformer
work,” Physiological Measurement, vol. 40, no. 5, article model with temporal features for ECG heartbeat classifica-
055002, 2019. tion,” IEEE International Conference on Bioinformatics and
[15] M. Salem, S. Taheri, and J. S. Yuan, “ECG arrhythmia classifi- Biomedicine, vol. 2019, pp. 898–905, 2019.
cation using transfer learning from 2-dimensional deep CNN [32] L. Wu, X. Xie, and Y. Wang, “ECG enhancement and R-peak
features,” in 2018 IEEE biomedical circuits and systems confer- detection based on window variabilityC//healthcare,” Multidis-
ence, pp. 1–4, 2018. ciplinary Digital Publishing Institute, vol. 9, no. 2, p. 227, 2021.
[16] E. Izci, M. A. Ozdemir, M. Degirmenci, and A. Akan, “Cardiac [33] E. H. Houssein, M. Hassaballah, I. E. Ibrahim, D. S.
arrhythmia detection from 2D ECG images by using deep AbdElminaam, and Y. M. Wazery, “An automatic arrhyth-
learning technique,” Medical Technologies Congress (TIPTE- mia classification model based on improved marine predators
KNO), vol. 2019, pp. 1–4, 2019. algorithm and convolutions neural networks,” Expert Systems
[17] J. Huang, B. Chen, B. Yao, and W. He, “ECG arrhythmia clas- with Applications, vol. 187, p. 115936, 2022.
sification using STFT-based spectrogram and convolutional [34] F. Li, Y. Xu, Z. Chen, and Z. Liu, “Automated heartbeat classi-
neural network,” IEEE Access, vol. 7, pp. 92871–92880, 2019. fication using 3-D inputs based on convolutional neural
Computational and Mathematical Methods in Medicine 13

network with multi-fields of view,” IEEE Access, vol. 7,

pp. 76295–76304, 2019.
[35] I. Goodfellow, J. Pouget-Abadie, M. Mirza et al., “Generative
adversarial nets,” Advances in Neural Information Processing
Systems, vol. 27, 2014.
[36] E. D. Cubuk, B. Zoph, D. Mane, V. Vasudevan, and Q. V. Le,
AutoAugment: Learning Augmentation Policies from Data,
Computer Vision and Pattern Recognition, 2018, http://arxi-
v.org/abs/1805.09501.
[37] K. Jiang, S. Liang, L. Meng, Y. Zhang, P. Wang, and W. Wang,
“A two-level attention-based sequence-to-sequence model for
accurate inter-patient arrhythmia detection,” IEEE Interna-
tional Conference on Bioinformatics and Biomedicine, pp.
1029–1033, 2020.
[38] S. Raj and K. C. Ray, “Sparse representation of ECG signals for
automated recognition of cardiac arrhythmias,” Expert Sys-
tems with Applications, vol. 105, pp. 49–64, 2018.
[39] A. Sellami and H. Hwang, “A robust deep convolutional neural
network with batch-weighted loss for heartbeat classification,”
Expert Systems with Applications, vol. 122, pp. 75–84, 2019.
[40] J. Niu, Y. Tang, Z. Sun, and W. Zhang, “Inter-patient ECG
classification with symbolic representations and multi-
perspective convolutional neural networks,” IEEE Journal of
Biomedical and Health Informatics, vol. 24, no. 5, pp. 1321–
1332, 2020.
[41] J. He, J. Rong, L. Sun, H. Wang, and Y. Zhang, “An advanced
two-step DNN-based framework for arrhythmia detection,” in
Pacific-Asia Conference on Knowledge Discovery and Data
Mining, pp. 422–434, Springer, Cham, 2020.
[42] F. M. Dias, H. L. Monteiro, T. W. Cabral, R. Naji, M. Kuehni,
and E. J. D. S. Luz, “Arrhythmia classification from single-lead
ECG signals using the inter-patient paradigm,” Computer
Methods and Programs in Biomedicine, vol. 202, p. 105948,
2021.
[43] Y. Kaya and H. Pehlivan, “Classification of premature ventric-
ular contraction in ECG,” International Journal of Advanced
Computer Science and Applications, vol. 6, no. 7, 2015.
[44] G. Sivapalan, K. Nundy, S. Dev, B. Cardiff, and J. Deepu,
“ANNet: a lightweight neural network for ECG anomaly
detection in IoT edge sensors,” IEEE Transactions on Biomed-
ical Circuits and Systems, p. 1, 2022.
[45] Y. Kaya, “Detection of bundle branch block using higher order
statistics and temporal features,” The International Arab Jour-
nal of Information Technology, vol. 18, no. 3, pp. 279–285,
2021.
[46] M. Ganeshkumar, V. Ravi, V. Sowmya, E. A. Gopalakrishnan,
and K. P. Soman, “Explainable deep learning-based approach
for multilabel classification of electrocardiogram,” IEEE
Transactions on Engineering Management, pp. 1–13, 2021.
[47] A. M. Shaker, M. Tantawi, H. A. Shedeed, and M. F. Tolba,
“Generalization of convolutional neural networks for ECG
classification using generative adversarial networks,” IEEE
Access, vol. 8, pp. 35592–35605, 2020.