Cardiac Arrhythmia Classiﬁcation in Electrocardiogram Signals with

Convolutional Neural Networks

Igor Lopes Souza

and Daniel Oliveira Dantas

Departamento de Computac¸

ao, Universidade Federal de Sergipe, S

ao Crist

ao, SE, Brazil

Keywords:

Electrocardiography, ECG, Atrial Fibrillation.

Abstract:

Electrocardiography is a frequently used examination technique for heart disease diagnosis. Electrocardiogra-

phy is essential in the clinical evaluation of patients who have heart disease. Through the electrocardiogram

(ECG), medical doctors can identify whether the cardiac muscle dysfunctions presented by the patient have

an inﬂammatory origin and early diagnosis of serious diseases that primarily affect the blood vessels and the

brain. The basis of arrhythmia diagnosis is the identiﬁcation of normal and abnormal heartbeats and their

classiﬁcation into different diagnoses based on ECG morphology. Heartbeats can be divided into ﬁve cate-

gories: non-ectopic, supraventricular ectopic, ventricular ectopic, fusion, and unknown beats. It is difﬁcult to

distinguish these heartbeats apart on the ECG as these signals are typically corrupted by outside noise. The

objective of this study is to develop a classiﬁer capable of classifying a patient’s ECG signals for the detec-

tion of arrhythmia in clinical patients. We developed a convolutional neural network (CNN) to identify ﬁve

categories of heartbeats in ECG signals. Our experiment was conducted with ECG signals obtained from a

publicly available MIT-BIH database. The number of instances was even out to ﬁve classes of heartbeats. The

proposed model achieved an accuracy of 99.33% and an F1-score of 99.44% in the classiﬁcation of ventricular

ectopic beats (VEB).

1 INTRODUCTION

According to the World Health Organization, cardio-

vascular diseases (CVDs) are the leading cause of

death in the world (McAloon et al., 2016). Arrhyth-

mia, a heart rhythm disorder, is considered one of the

most common disorders of the heart. Arrhythmia is

a problem with the rate or rhythm of the heartbeat.

During an arrhythmia, the heart may beat too fast, too

slow, or with an irregular rhythm. Atrial ﬁbrillation

(AF) is the most prevalent case of arrhythmia. AF

causes irregular heartbeats. In AF, the electrical ac-

tivity of the atria (the heart’s upper chambers) is ir-

regular, inconsistent, and not synchronized with ven-

tricles (Hagiwara et al., 2018).

AF is diagnosed by interpreting the ECG. Au-

tomatic diagnosis is useful in home settings, where

an ECG interpretation specialist is not available to

diagnose AF (Mant et al., 2007). Classiﬁcation of

ECG signals is necessary for the automatic diagno-

sis of arrhythmia. To improve AF detection, machine

learning methods were used by various authors (Lown

et al., 2020; Pollock et al., 2020; Shoemaker et al.,

https://orcid.org/0000-0002-2499-4607

https://orcid.org/0000-0002-0142-891X

2020). Recently, Sanchez successfully experimented

with the latest and most innovative convolutional neu-

ral networks (CNN) (S

anchez and Cervera, 2019).

Deep convolutional neural networks have the capa-

bility of hierarchical feature learning, which allows

the neural network to differentiate and generalize

ECG signal patterns with higher accuracy than an ex-

pert (Chen et al., 2022; Kiranyaz et al., 2021). CNNs

have been used to diagnose arrhythmias, and coro-

nary artery diseases, and classify strokes (Zhiqiang

and Jun, 2017).

Many approaches to arrhythmia heartbeat classi-

ﬁcation with CNN have been proposed. Han (Han

and Shi, 2020) presents a method to detect and local-

ize myocardial infarction by combining a multiple-

lead residual neural network (ML-ResNet) frame-

work with three residual blocks and feature fusion us-

ing 12-lead ECG recordings.

Qiyang Xie (Xie et al., 2021) used ResNet34 to

train a model with the morphological characteristics

of ECG signals and obtain meaningful information

from ECG signals.

Xiong (Xiong et al., 2017) proposed a purely data-

driven, deep learning pipeline, a 16-layer CNN, for

the automatic classiﬁcation of ECG signals from the

356

Souza, I. and Dantas, D.

Cardiac Arrhythmia Classiﬁcation in Electrocardiogram Signals with Convolutional Neural Networks.

DOI: 10.5220/0011682800003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 356-362

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Computing in Cardiology (CinC) Challenge 2017 into

four categories, including AF. The large dataset of

ECG data recorded from patients and labeled by ex-

perts provided a framework for developing and vali-

dating their approach to ECG diagnosis.

Zhi Li (Li et al., 2020) developed a deep learning

method for cardiac arrhythmia classiﬁcation based on

ResNet. The design consists of a 1D 31-layer con-

volutional residual network. The algorithm includes

four residual blocks, each of which consists of three

layers of 1D convolutions, three layers of batch nor-

malization (BN), three layers of the rectiﬁed linear

unit (ReLU) activation function, and a structure of

identity shortcut connections. The 2-lead ECG sig-

nals were used in combination with deep learning

techniques to automatically identify the normal, left

group, right group, premature atrial, and premature

ventricular contraction heartbeats.

Zhu (Zhu et al., 2020) developed a deep learn-

ing approach for the automated diagnosis of multiple

cardiac rhythm labels or conduction abnormalities by

real-time ECG signal analysis. The dataset used was

obtained from ECG data with 10s in length and 12-

channel format. The data is from adult patients, with

21 distinct rhythm classes for the diagnosis of simul-

taneous cardiac arrhythmias, i.e., patients with multi-

ple heart diseases.

Kiranyaz (Kiranyaz et al., 2017) proposed a per-

sonalized health monitoring system that can detect

early occurrences of arrhythmias from a patient’s

ECG signal by modeling common causes of arrhyth-

mias in the signal domain as degradation from nor-

mal ECG beats to abnormal beats. Using the degra-

dation models, abnormal beats were created from the

patient’s average normal beat. A simple 1D convolu-

tional neural network was trained using real normal

beats and synthesized abnormal beats.

Han (Han and Shi, 2020) presented a method to

detect and localize myocardial infarction by com-

bining an ML-ResNet framework with three residual

blocks and feature fusion using 12-lead ECG record-

ings. The single-lead feature branching network is

trained to automatically learn local features of dif-

ferent levels between different layers, which can be

used to characterize the spatial representation of the

ECG. The main features are merged as global fea-

tures. For the generalization and evaluation of the

proposed method in clinics, intra-patient and inter-

patient schemes were used.

Acharya (Acharya et al., 2017) developed a nine-

layer deep convolutional neural network to identify

ﬁve different categories of heartbeats in ECG signals:

non-ectopic beat, supraventricular ectopic beat, ven-

tricular ectopic beat, fusion beat, and unknown beat.

The experiment was conducted on noise-attenuated

and non-attenuated data sets from a public database,

MIT-BIH. This set was artiﬁcially augmented to equal

the number of instances of the ﬁve heartbeat classes

and ﬁltered to remove high-frequency noise.

Xiang (Xiang et al., 2018) proposed an accu-

rate method for patient-speciﬁc ECG beat classiﬁca-

tion, which adopts morphological features and tim-

ing information. As to the morphological features

of a heartbeat, attention-based two-level 1-D CNN

is incorporated in the proposed method to extract

different-grained features automatically by focusing

on various parts of a heartbeat. The timing informa-

tion, the difference between previous and post RR in-

tervals, is computed as a dynamic feature. Both the

extracted morphological features and the interval dif-

ference are used by multi-layer perceptron (MLP) for

classifying ECG signals.

Ali Sellami (Sellami and Hwang, 2019) proposed

a new type of deep convolutional neural network for

heartbeat classiﬁcation. A batch-weighted loss func-

tion was created to quantify the loss and decrease

the imbalance between classes. The loss weights

change dynamically as the distribution of classes in

each batch changes.

Schwab (Schwab et al., 2017) proposed a machine

learning approach based on recurrent neural networks

(RNN) to analyze different cardiac arrhythmias with

only a single lead and short ECG recordings, below

10s. To facilitate training dependencies on the tem-

poral dimension, a new task formulation was intro-

duced that takes advantage of the natural beat-based

segmentation of ECG signals.

Rahhal (Al Rahhal et al., 2019) proposed a novel

end-to-end architecture based on a dense convolu-

tional network (DCN) for ECG signal classiﬁcation.

The architecture is based on two main modules: the

ﬁrst is a generative module and the second is a dis-

criminative module. The generative module con-

verts the one-dimensional ECG signal into an image

through fully connected up-sampling layers and con-

volutional layers. The discriminative module receives

the image from the generative module and performs

feature learning and classiﬁcation.

Zhai (Zhai and Tin, 2018) developed a high-

performance ECG-based arrhythmic beat classiﬁca-

tion system. The classiﬁer was designed based on a

CNN. The single-channel ECG signal was segmented

into heartbeats according to the change in beats. Zhai

provided accurate ECG classiﬁcation tools.

We believe that it is possible to further improve

the accuracy, sensitivity, speciﬁcity, precision, and

F1-score of CNN heartbeat classiﬁers. Our study

aims to improve the classiﬁcation metrics by increas-

Cardiac Arrhythmia Classiﬁcation in Electrocardiogram Signals with Convolutional Neural Networks

357

ing the number of convolution layers in the couple-

convolution implementation and using a new archi-

tecture with triple-convolution (Uchida et al., 2018)

for the classiﬁer in conjunction with ﬁne-tuning for

further optimization. The classiﬁcation of arrhyth-

mia signals in public ECG datasets will generate more

precise and accurate results. The improved classiﬁca-

tion of ECG signals will generate more accurate re-

sponses in the detection of cardiac arrhythmias, fa-

cilitating the health care of patients. Our neural net-

work architecture was inspired by Xiong architecture

design (Xiong et al., 2017), but with a faster architec-

ture, fewer loops, more convolutions for feature ex-

traction, and a higher resulting F1-Score.

2 METHODOLOGY

In this study, we created a classiﬁer capable of distin-

guishing the different types of heartbeats and detect-

ing cardiac arrhythmia. This architecture was ﬁne-

tuned so that the models achieved the highest valida-

tion accuracy and F1-score possible. Our model was

evaluated in the test set. The ECG heartbeat classiﬁer

is composed of two main steps: preprocessing and

classiﬁcation. The network architecture is shown in

Figure 1. The implementation of this methodology is

publicly available

and was coded in Python using

Tensorﬂow, Keras, and Numpy.

2.1 Dataset

In this study, we used the ECG Heartbeat Categoriza-

tion Dataset, freely available in the Internet

. We

used only the portion of the dataset derived from the

Physio Bank MIT-BIH Arrhythmia database (Mark

and Moody, 1988). This database consists of a 48

half-hour long ECG recordings from 47 subjects—

obtained with a Lead II ECG conﬁguration—that was

band-pass ﬁltered over the frequency range from 0.1

to 100Hz and digitized at 360 samples per second.

Furthermore, these recordings were interpreted and

validated by at least two cardiologists. The database

consists of annotations for both heartbeat class in-

formation and R-peak position information veriﬁed

by two or more expert cardiologists. The 17 beat

types can be grouped into ﬁve beat classes deﬁned by

the Association of Advancement for Medical Instru-

mentation (AAMI) which follows the American Na-

tional Standard for Ambulatory ECGs (ANSI/AAMI

EC38:2007) recommendations.

https://github.com/Igor-Lopes-Souza/VISAPP-2023

https://www.kaggle.com/datasets/shayanfazeli/heartbeat

Figure 1: Neural network architecture.

Figure 2: Triple-convolution.

2.2 Preprocessing

The raw MIT-BIH signal is corrupted by myoelectric

interference, power line interference, and line drift.

To remove these noises, the raw ECG signal is ﬁl-

tered using wavelet ﬁlters. The raw signal is decom-

posed by Daubechies wavelet 6 (db6) at six levels, and

wavelet coefﬁcients from the third to the sixth level

were retained and used for signal reconstruction (Shi

et al., 2019). After noise removal, we segmented the

signal for heartbeats by taking advantage of informa-

tion from the positions of the R-peaks annotated in the

MIT-BIH arrhythmia database. Each heartbeat con-

sists of 300 samples: 149 before and 150 after the

R-peak position.

2.3 Classiﬁer Architecture

Figure 1 shows the schematic of our CNN classi-

ﬁer. The network is composed of convolutional lay-

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

358

Table 1: Hyperparameter values chosen in classiﬁer ﬁne-tuning.

Parameters Values Chosen Value

Dropout 0.10, 0.25, 0.30, 0.50 0.50

Optimizer Adam, Adamax, SGD Adam

Activation function

Relu, Selu, Elu, Softmax, Softplus Relu

Batch size

100, 250, 500, 1000, 1500 500

Epochs

10, 25, 75, 125, 175, 300, 1000 75

Loss function

Binary cross-entropy, Categorical cross-entropy,

Poisson, Kullback-Leibler divergence, Huber

Categorical cross-entropy

Learning rate

0.01, 0.001, 0.0001, 0.00001 0.0001

β1

0.900, 0.990, 0.999 0.900

β2

0.900, 0.990, 0.999 0.990

Table 2: Comparison of the proposed algorithm classiﬁcation using ventricular ectopic beats (VEB).

ACC SEN SPE PRE F1S

Martis (Martis et al., 2014)

99.45% 99.61% 99.99% 99.99% 99.8%

Proposed classiﬁer

99.33% 99.59% 99.30% 99.12% 99.44%

Sellami (Sellami and Hwang, 2019)

99.48% 96.97% 99.87% 98.83% 97.80%

Acharya (Acharya et al., 2017)

94.03% 96.71% 91.54% 97.85% 97.27%

Zhai (Zhai and Tin, 2018)

99.10% 96.40% 99.50% 96.40% 96.40%

Yande (Xiang et al., 2018)

99.20% 93.70% 99.60% 94.80% 94.20%

Jiang (Jiang and Kong, 2007)

98.80% 94.30% 99.40% 95.30% 94.70%

Ince (Ince et al., 2009)

97.60% 83.60% 98.10% 87.40% 85.40%

ers, subsampling layers, fully connected layers, and

a softmax layer. The convolutional layers perform

the convolution operations on the output of a previ-

ous layer using the current convolution kernel (ω

The merge layer adds two layers, in our case the sec-

ond convolution layer and the ﬁrst activation function

of each execution. Usually, each convolution layer

is followed by a subsampling layer. However, to fa-

cilitate mapping between the heartbeat category and

its waveform, we use a triple-convolution structure

to achieve a more powerful ﬁtting capability (Uchida

et al., 2018). Figure 2 shows the structure of a triple-

convolution layer sequence.

The subsampling layer was used to reduce by half

the input size of the next layer, compressing the size

of the ECG data, reducing the number of computa-

tions and extracting useful features, our max pooling

size is set to 5 with a stride of 2 in all pooling lay-

ers. The function max-pooling was used to obtain

the maximum value inside a region around each posi-

tion in the input matrix (Murray and Perronnin, 2014).

Fully connected layers were used to increase the num-

ber of nonlinear operations (Xu et al., 2019).

In this study, we use the ReLu function as an

activation function in both convolutional layers and

fully connected layers (Nair and Hinton, 2010; Girosi

et al., 1995). In the output layer, we use the acti-

vation function softmax to obtain the ﬁve heartbeat

categories (non-ectopic beat, supraventricular ectopic

beat, ventricular ectopic beat, fusion beat, and un-

known beat) (Nwankpa et al., 2018).

2.4 Training Method

The goal of training is to reduce the value of the loss

function L, i.e., to decrease the model loss and ad-

just the weights and biases so that Equation 1 ﬁts the

model training set. The cross-entropy function is used

as the loss function (Xu and Liu, 2020):

We update the weights and offsets using the Adam

optimizer (Kingma and Ba, 2014). First, a batch

of samples was sent to calculate the gradient of the

Equation 1, and we set the batch size to 256:

g =

∇

∑

L( f (x

(i)

;θ), y

(i))

. (1)

The g is the gradient value, m is the batch size, θ is

the parameter to be updated, f (x

(i)

;θ) is the heartbeat

type predicted by the i-th sample, y

(i)

is the actual type

of the i-th sample, and L is the loss function. The

and v

represent the ﬁrst and second estimates of

the moment of the gradient. The ˆm

and ˆv

are the

corresponding bias corrections. The β

and β

are the

decay rates for the moment estimates, set to 0.900 and

0.990.

The regularization dropout (Hinton et al., 2012;

Srivastava et al., 2014) was used to avoid overﬁtting

Cardiac Arrhythmia Classiﬁcation in Electrocardiogram Signals with Convolutional Neural Networks

359

Table 3: Comparison of proposed implementations.

ACC SEN SPE PRE F1S

With

preprocessing

With

subsampling

Triple

convolution

99.33% 99.59% 99.30% 99.12% 99.44%

Simple

convolution

95.32% 95.73% 98.83% 96.39% 95.40%

Without

subsampling

Triple

convolution

95.40% 95.27% 94.70% 95.25% 95.35%

Simple

convolution

90.45% 90.14% 92.83% 90.89% 90.44%

Without

preprocessing

With

subsampling

Triple

convolution

89.65% 89.00% 97.41% 91.63% 89.50%

Simple

convolution

87.85% 83.25% 96.06% 90.63% 87.56%

Without

subsampling

Triple

convolution

86.68% 88.69% 96.80% 90.05% 86.67%

Simple

convolution

87.20% 90.40% 90.50% 89.20% 85.80%

and excessive specialization in the training dataset,

in the convolutional layer, and in the fully connected

layers. The dropout allows the weights of the hidden

layer neurons to be randomly set to zero during train-

ing, causing these nodes to be ignored. After deﬁning

the architecture, ﬁne-tuning was performed to obtain

the best number of epochs. The hyperparameters that

decreased the classiﬁer training time and increased

accuracy and F1-score were obtained and displayed

in Table 1.

Since the best optimization method was obtained

with the Adam function, we also needed to optimize

the learning rate, β1 and β2 values. We tried the val-

ues 0.01, 0.001, 0.0001 and 0.00001 for the learning

rate. Furthermore, to obtain a variable learning rate,

we tested β1 and β2 with the values 0.900, 0.990 and

0.999. The best result was obtained with a 0.0001

learning rate, 0.900 β1 and 0.990 β2. We tested sev-

eral other loss functions to optimize the classiﬁer, and

the categorical cross-entropy was the one that gen-

erated the best results. We needed to ﬁnd the min-

imum number of epochs necessary to maximize the

accuracy. Excessive training could cause overﬁtting

and incapacity to generalize and evaluate new images.

We started the test with 100 epochs and increased this

value until 1500 epochs. This test showed that for 500

epochs or more, the accuracy remained stable.

3 RESULTS AND DISCUSSION

We performed classiﬁcation experiments on 44

recordings from the MIT-BIH arrhythmia database,

among the 48 recordings obtained from 47 patients

studied by the BIH arrhythmia laboratory, and the

heartbeats were classiﬁed according to the recom-

mendation of the AAMI.

The training dataset contains a total of 375 rep-

resentative beats, including 75 from each class: type-

N, non-ectopic beats; type-S, supraventricular ectopic

beats; type-V, ventricular ectopic beats; type-F, fusion

beats and type-Q, unknown beats. The representative

beats are randomly sampled from each class of the

ﬁrst 20 recordings (chosen in the range of 100 to 124)

from the MIT-BIH database. The neural networks are

trained with a total of 245 common training beats, and

a variable number of beats depending on the patient’s

heart rate, so less than 1% of the total beats are used

for training. The 24 unused recordings are used as test

patterns for performance evaluation.

Classiﬁcation performance is measured using the

statistical error metrics found in the literature (Chen

et al., 2022): accuracy (ACC), sensitivity (SEN),

speciﬁcity (SPE), precision (PRE), and F1-score

(F1S). The F1-score measures the overall perfor-

mance of the beat classiﬁcation, as shown in Table 2.

Our model was implemented using the Tensor-

Flow framework. The training time of each epoch

was approximately 5s, and the maximum epoch num-

ber was set to 75. Table 2 shows that the implemented

classiﬁcation algorithm has an F1-score value compa-

rable to those of other studies obtaining better results,

presenting the second best in Table 2. We show only

the VEB in the comparison as it is the most commonly

used among other studies. Table 3 shows the results

of different convolutional architectures. In this study,

our best model achieved an accuracy of 99.33%, sen-

sitivity of 99.59%, speciﬁcity of 99.30%, precision of

99.12% and F1-score of 99.44%.

Figure 3 shows the confusion matrix of the clas-

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

360

Figure 3: Confusion matrix for heartbeat classiﬁcation on

the test set.

siﬁcation results of the test set. The model is able

to make accurate predictions and distinguish different

classes. The main reason behind this might be the fact

that we have used ﬁne-tuning to optimize our model,

which allows us to better train our classiﬁer.

4 CONCLUSIONS

In this study, we designed an ECG signals classiﬁer

for cardiac arrhythmia detection using CNNs. The

proposed model achieved an accuracy of 99.33% and

an F1-score of 99.44% in the classiﬁcation of ven-

tricular ectopic beats (VEB). In order to optimize our

model, we ﬁne-tuned our variables and functions, the

selected values compose our ﬁnal version of the clas-

siﬁer and are displayed in Table 1. Compared with the

methods in previous literature, our model performed

better in terms of VEB classiﬁcation accuracy, and

F1-score.

The referenced authors in Table 2 achieved high

accuracy, sensitivity, speciﬁcity, precision and F1-

score with private datasets. On the other hand, our

study managed to obtain high results for these met-

rics with a public dataset. Our trained CNN heart-

beat classiﬁer model can be used for real-life and real-

time applications. It can also be used to analyze other

biosignals by changing the training the dataset and

input size before use. Future work may reﬁne this

approach with a better set of hyperparameter values

and different augmentation strategies. An F1-score of

99.00% is accurate enough for cardiovascular disease

detection in home devices. The method has the poten-

tial to be adapted to analyze other biosignals.

REFERENCES

Acharya, U. R., Oh, S. L., Hagiwara, Y., Tan, J. H., Adam,

M., Gertych, A., and San Tan, R. (2017). A deep

convolutional neural network model to classify heart-

beats. Computers in biology and medicine, 89:389–

396.

Al Rahhal, M. M., Bazi, Y., Almubarak, H., Alajlan, N.,

and Al Zuair, M. (2019). Dense convolutional net-

works with focal loss and image generation for elec-

trocardiogram classiﬁcation. IEEE Access, 7:182225–

182237.

Chen, S. W., Wang, S. L., Qi, X. Z., Samuri, S. M., and

Yang, C. (2022). Review of ECG detection and clas-

siﬁcation based on deep learning: Coherent taxon-

omy, motivation, open challenges and recommenda-

tions. Biomedical Signal Processing and Control,

74:103493.

Girosi, F., Jones, M., and Poggio, T. (1995). Regulariza-

tion theory and neural networks architectures. Neural

computation, 7(2):219–269.

Hagiwara, Y., Fujita, H., Oh, S. L., Tan, J. H., San Tan, R.,

Ciaccio, E. J., and Acharya, U. R. (2018). Computer-

aided diagnosis of atrial ﬁbrillation based on ecg sig-

nals: A review. Information Sciences, 467:99–114.

Han, C. and Shi, L. (2020). Ml–resnet: A novel network to

detect and locate myocardial infarction using 12 leads

ecg. Computer methods and programs in biomedicine,

185:105138.

Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. R. (2012). Improving neural

networks by preventing co-adaptation of feature de-

tectors. arXiv preprint arXiv:1207.0580.

Ince, T., Kiranyaz, S., and Gabbouj, M. (2009). A generic

and robust system for automated patient-speciﬁc clas-

siﬁcation of ECG signals. IEEE Transactions on

Biomedical Engineering, 56(5):1415–1426.

Jiang, W. and Kong, S. G. (2007). Block-based neural

networks for personalized ECG signal classiﬁcation.

IEEE Transactions on Neural Networks, 18(6):1750–

1761.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. arXiv preprint

arXiv:1412.6980.

Kiranyaz, S., Avci, O., Abdeljaber, O., Ince, T., Gabbouj,

M., and Inman, D. J. (2021). 1d convolutional neu-

ral networks and applications: A survey. Mechanical

systems and signal processing, 151:107398.

Kiranyaz, S., Ince, T., and Gabbouj, M. (2017). Personal-

ized monitoring and advance warning system for car-

diac arrhythmias. Scientiﬁc reports, 7(1):1–8.

Li, Z., Zhou, D., Wan, L., Li, J., and Mou, W. (2020). Heart-

beat classiﬁcation using deep residual convolutional

neural network from 2-lead electrocardiogram. Jour-

nal of Electrocardiology, 58:105–112.

Lown, M., Brown, M., Brown, C., Yue, A. M., Shah, B. N.,

Corbett, S. J., Lewith, G., Stuart, B., Moore, M., and

Little, P. (2020). Machine learning detection of atrial

ﬁbrillation using wearable technology. PLoS One,

15(1):e0227401.

Cardiac Arrhythmia Classiﬁcation in Electrocardiogram Signals with Convolutional Neural Networks

361

Mant, J., Fitzmaurice, D. A., Hobbs, F. R., Jowett, S., Mur-

ray, E. T., Holder, R., Davies, M., and Lip, G. Y.

(2007). Accuracy of diagnosing atrial ﬁbrillation on

electrocardiogram by primary care practitioners and

interpretative diagnostic software: analysis of data

from screening for atrial ﬁbrillation in the elderly

(safe) trial. Bmj, 335(7616):380.

Mark, R. and Moody, G. (1988). Mit-bih arrhythmia

database directory. Cambridge: Massachusetts Insti-

tute of Technology.

Martis, R. J., Acharya, U. R., Adeli, H., Prasad, H., Tan,

J. H., Chua, K. C., Too, C. L., Yeo, S. W. J., and Tong,

L. (2014). Computer aided diagnosis of atrial arrhyth-

mia using dimensionality reduction methods on trans-

form domain representation. Biomedical signal pro-

cessing and control, 13:295–305.

McAloon, C. J., Boylan, L. M., Hamborg, T., Stallard, N.,

Osman, F., Lim, P. B., and Hayat, S. A. (2016). The

changing face of cardiovascular disease 2000–2012:

An analysis of the world health organisation global

health estimates data. International journal of car-

diology, 224:256–264.

Murray, N. and Perronnin, F. (2014). Generalized max

pooling. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 2473–

2480.

Nair, V. and Hinton, G. E. (2010). Rectiﬁed linear units

improve restricted boltzmann machines. In Icml.

Nwankpa, C., Ijomah, W., Gachagan, A., and Marshall, S.

(2018). Activation functions: Comparison of trends in

practice and research for deep learning. arXiv preprint

arXiv:1811.03378.

Pollock, K. G., Sekelj, S., Johnston, E., Sandler, B., Hill,

N. R., Ng, F. S., Khan, S., Nassar, A., and Farooqui, U.

(2020). Application of a machine learning algorithm

for detection of atrial ﬁbrillation in secondary care.

IJC Heart & Vasculature, 31:100674.

anchez, F. R. and Cervera, J. G. (2019). ECG classiﬁcation

using artiﬁcial neural networks. In Journal of Physics:

Conference Series, volume 1221, page 012062. IOP

Publishing.

Schwab, P., Scebba, G. C., Zhang, J., Delai, M., and Karlen,

W. (2017). Beat by beat: Classifying cardiac arrhyth-

mias with recurrent neural networks. In 2017 Com-

puting in Cardiology (CinC), pages 1–4. IEEE.

Sellami, A. and Hwang, H. (2019). A robust deep convo-

lutional neural network with batch-weighted loss for

heartbeat classiﬁcation. Expert Systems with Applica-

tions, 122:75–84.

Shi, H., Wang, H., Zhang, F., Huang, Y., Zhao, L., and Liu,

C. (2019). Inter-patient heartbeat classiﬁcation based

on region feature extraction and ensemble classiﬁer.

Biomedical Signal Processing and Control, 51:97–

105.

Shoemaker, M. B., Shah, R. L., Roden, D. M., and Perez,

M. V. (2020). How will genetics inform the clini-

cal care of atrial ﬁbrillation? Circulation research,

127(1):111–127.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. (2014). Dropout: a simple way

to prevent neural networks from overﬁtting. The jour-

nal of machine learning research, 15(1):1929–1958.

Uchida, K., Tanaka, M., and Okutomi, M. (2018). Cou-

pled convolution layer for convolutional neural net-

work. Neural Networks, 105:197–205.

Xiang, Y., Luo, J., Zhu, T., Wang, S., Xiang, X., and Meng,

J. (2018). ECG-based heartbeat classiﬁcation using

two-level convolutional neural network and rr interval

difference. IEICE TRANSACTIONS on Information

and Systems, 101(4):1189–1198.

Xie, Q., Wang, X., Sun, H., Zhang, Y., and Lu, X.

(2021). ECG signal detection and classiﬁcation of

heart rhythm diseases based on ResNet and LSTM.

Advances in Mathematical Physics, 2021.

Xiong, Z., Stiles, M. K., and Zhao, J. (2017). Robust ecg

signal classiﬁcation for detection of atrial ﬁbrillation

using a novel neural network. In 2017 Computing in

Cardiology (CinC), pages 1–4. IEEE.

Xu, Q., Zhang, M., Gu, Z., and Pan, G. (2019). Over-

ﬁtting remedy by sparsifying regularization on fully-

connected layers of cnns. Neurocomputing, 328:69–

74.

Xu, X. and Liu, H. (2020). ECG heartbeat classiﬁcation

using convolutional neural networks. IEEE Access,

8:8614–8619.

Zhai, X. and Tin, C. (2018). Automated ECG classiﬁcation

using dual heartbeat coupling based on convolutional

neural network. IEEE Access, 6:27465–27472.

Zhiqiang, W. and Jun, L. (2017). A review of object de-

tection based on convolutional neural network. In

2017 36th Chinese control conference (CCC), pages

11104–11109. IEEE.

Zhu, H., Cheng, C., Yin, H., Li, X., Zuo, P., Ding, J.,

Lin, F., Wang, J., Zhou, B., Li, Y., et al. (2020).

Automatic multilabel electrocardiogram diagnosis of

heart rhythm or conduction abnormalities with deep

learning: a cohort study. The Lancet Digital Health,

2(7):e348–e357.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

362