Malleswari et al.

EURASIP Journal on Advances

EURASIP Journal on Advances in Signal Processing (2024) 2024:104
https://doi.org/10.1186/s13634-024-01197-1 in Signal Processing

RESEARCH Open Access

Deep learning‑assisted arrhythmia

classification using 2‑D ECG spectrograms
Pinjala N Malleswari1*, Venkata krishna Odugu2†, T. J. V. Subrahmanyeswara Rao3† and T. V. N. L. Aswini4†

†
Venkata krishna Odugu, T. J. V.
Subrahmanyeswara Rao, and T. Abstract
V. N. L. Aswini have contributed This article studies modern classification techniques in ECG signals through the transfer
equally to this work.
learning approach with CNN (Convolutional Neural Network). The proposed pre-
*Correspondence: trained network combines an Imagenet with huge labeled image datasets and a sepa-
[email protected]
1
rate network composed of fully connected layers. This method uses the CWT (Con-
Electronics and Communication
Technology, Sasi Institute
tinuous Wavelet Transform) to construct a time-frequency visualization of ECG signals,
of Technology & Engineering, which are subsequently transformed into RGB images. The developed images are
Tadepalligudem, AP 534101, plugged into a pre-trained CNN to retrieve the desired features. We next employ
India
2
Electronics and Communication
supervised learning to train the neural network on the ECG labeled data using CNN
Engineering, CVR College features. To train a Deep Neural Network, three sets of PhysioNet databases are used:
of Engineering, Hyderabad, MIT-BIH (ARR) Arrhythmia, NSR (Normal Sinus Rhythm), and BIDMC CHF (Congestive
Telangana 501510, India
3
Electronics and Communication
Heart Failure). The classification Accuracy, Sensitivity, Specificity, F1-score, Precision,
Engineering, Sasi Institute and Detection Error Rate of the CNN classifier are compared to AlexNet, GoogleNet,
of Technology & Engineering, Vgg16, and SqueezeNet pre-trained networks. Among all these networks, SqueezeNet
Tadepalligudem, AP 534101,
India
provides an Acc of 98.7%, Se of 99.1%, Sp of 99.20%, F1-score of 98.33%, Precision
4
Electronics and Communication of 98.67%, and DER of 0.89%. For further investigation, the technique suggested can be
Engineering, Sri Vasavi implemented in addition to Bi-LSTM on some real ECG data.
Engineering College,
Tadepalligudem, AP 534101, Keywords: Electrocardiogram, Convolutional Neural Networks, Classification, Deep
India learning

1 Introduction
The heartbeat, a fundamental physiological signal, is a crucial indicator of cardiac activ-
ity, reflecting various anomalies influenced by age and lifestyle. Deep learning, pivotal
in modern computer-based solutions, facilitates automated feature detection from
input data. It enables information extraction, prediction, and intelligent decision-
making, leveraging training datasets. Heart irregularities are increasing, posing global
health concerns. ECG signals remain key for detecting and assessing heart abnormali-
ties. Sedentary lifestyles contribute to rising heart and cancer issues. Xu [1] describes
an IoT-assisted ECG tracking system that ensures secure data transfer for healthcare. It
employs IoT devices for remote ECG data collection, securely transmitting it to health-
care providers.
The authors [2, 3] addressed an efficient classification method using Convolutional
Neural Networks (CNN) by monitoring healthy heart activity via ECG analyzing P, QRS,

© The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0
International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you
modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of
it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise
in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy
of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/.
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 2 of 15

T waves, and QT intervals. Shaker et al. [4] introduced a new data augmentation method
with generative adversarial network (GAN) to balance the dataset as well as detect heart
rate-affecting conditions. The authors applied two deep CNN methodologies from end
to end and two-stage model. This method integrates feature extraction, reduction, and
classification of heart beats. Hsu and Liu [5] implemented a new strategy to fold ECG
signals into fingerprint format, which could have implications for efficient storage, trans-
mission, and analysis of ECG data, potentially advancing the diagnostics of cardiovascu-
lar and healthcare technology.
Wasimuddin et al. [6] presented a review on ECG signal analysis using machine learn-
ing approaches like neural networks and support vector machines (SVMs). Improper
signals can lead to stroke or heart failure. NSR signals a healthy heart, while CHF
indicates compromised functionality due to inadequate circulation. To overcome this
problem, Elgendi and Menon [7] presented a novel deep learning (DL) algorithm by
combining CNN and bi-directional long short-term memory (Bi-LSTM). This algorithm
not only increases the classification accuracy but also reduces the training time. Gao [8]
presented a study on DL models such as CNN, RNN, LSTM, DBN, etc. for the identifica-
tion of irregular heartbeats. Many researchers have recommended a variety of DL mod-
els to increase the accuracy of multiple learning tasks, including multilayer perception
(MLP), CNN, RNN, long short-term memory (LSTM), and deep belief network (DBN)
[9]. Yildirim et al. [10] projected a novel algorithm for the classification of cardiac disor-
ders, which enhances the signal quality. A deep CNN is employed to obtain the features
and then fed into an LSTM network, which process the sequential data. Incorporating
deep coded features with LSTM networks creates an effective strategy for arrhythmia
classification, possibly enhancing the accuracy and efficacy of automated ECG monitor-
ing in clinical environments.
A CNN method with improved accuracy for automatically detecting regular and myo-
cardial infarction ECG beats in both noisy and noise-free conditions is investigated by
[11]. This method attained an average accuracy of 93.53% and 95.22% for ECG beats with
and without noise reduction, respectively. Furthermore, this study did not involve any
feature extraction or selection. As a result, this method can recognize unfamiliar ECG
signals with high accuracy even in noisy conditions. Malleswari et al. [12] implemented
a pre-trained neural network to analyze cardiac severity classification. The classification
of cardiac severity is accomplished through the utilization of existing models. In order
to enhance the efficacy and accuracy of cardiac diagnostics, pre-trained networks are
implemented.
Ting [13] developed an algorithm for the detection of AF (atrial fibrillation) using ECG
signals. This technique uses machine learning to accurately identify AF with high accu-
racy and reliability. Hannun [14] introduced a deep neural network to train an ambula-
tory ECG records, employing CNN for the extraction and classification of features. A
spectrogram is created by converting a one-dimensional (1-D) ECG waveform into two-
dimensional (2-D) signals using short time Fourier transform (STFT), which is then used
as a CNN classifier. Amin et al. [15] study proposes a 2D CNN model for classifying
ECG signals into eight categories, including arrhythmias. Initially, 2D spectrograms are
obtained by applying STFT on 1D ECG signals. The presented CNN architecture con-
sists of four convolutional and pooling layers for feature extraction and classification,
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 3 of 15

which provides high sensitivity and specificity, indicating its effectiveness in arrhythmia
classification.
Hong [16] investigated the usage of deep learning in ECG data processing with an
emphasis on CNN and RNN for the detection and classification of arrhythmia. It
addresses new methodologies such as transfer learning and federated learning. Gaddam
[17] presented a transferred DL algorithm to classify the ECG signal into four classes
such as ARR, NSR, Malignant Ventricular Ectopy (MVE), and CHF. In this method,
2D scalogram images are trained using CNN with pre-trained AlexNet with an accu-
racy of 95.67%. To enhance the classification accuracy with CNN, Jia et al. [18] fused
the structure and tempo of heartbeats to get a 2D information vector; then, it was pro-
cessed using a flexible learning rate and weighted dropout techniques. Malleswari et al.
[19] presented a novel algorithm that optimizes the detection of QRS complexes in ECG
signals by integrating DWT and EMD. In experiments conducted on a variety of ECG
datasets, this hybrid approach outperforms standard techniques concerning reliability as
well as accuracy.
Rahman et al. [20] implemented a deep CNN with transfer learning strategy to clas-
sify heartbeats and assessed it on PhysioNet’s MIT-BIH and PTB Diagnostics datasets.
Novel deep learning algorithms are developed based on CNN with bidirectional LSTM
showing more accuracy and less training time [21]. Panda et al. [22] introduced a fixed
frequency range empirical wavelet transform (FFREWT) deep CNN for the detection of
SVCA (Shockable Ventricular Cardiac Arrhythmias), and the obtained average accuracy
is 97.592%. To enhance the classification accuracy, Kanani and Padole [23] offer a modi-
fied DL architecture with stability. To extract the QRS complex, a DTCWT (dual-tree
complex wavelet transform) feature extraction method is applied, and its performance is
compared to that of the DWT using a multi-layer BPNN by [24]. A mobile device-opti-
mized deep neural network (DNN) for ECG signal classification is introduced by [25].
The DNN correctly classifies signals with minimum processing cost on ECG datasets,
allowing mobile cardiac monitoring in real time.
In this work, we transformed 1D ECG signals into 2D images and used CWT to
extract time-frequency components, which were then fed to pre-trained AlexNet, Goog-
leNet, Vgg16, and SqueezeNet CNN classifiers. Finally, their performance measures are
compared with existing methods and this paper uses the MIT-BIT arrhythmia database
as the data for the experiments in the following chapters.

2 Materials and methods

The projected system involves three stages: generating ECG signal time-frequency com-
ponents, converting 1D signals into 2D images, and using pre-trained networks for
transfer learning and CNN classification. The CWT-CNN classifier is compared with
Alexnet, GoogleNet, Vgg16, and Squeeze net models in terms of Acc, Se, Sp, Precision,
F1-score, and DER. The proposed pre-trained CNN architecture is shown in Fig. 1,
which consists of ten 2D convolutional layers and each layer is followed by a pooling
layer. To provide the ultimate classification, a softmax layer of three neurons is employed
as the output. A fully connected layer works across the last layer of pooling and the out-
put layer to represent the attributes learnt by the CNN algorithm.
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 4 of 15

Fig. 1 Block diagram of our pre-trained CNN model

Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 5 of 15

2.1 Proposed methodology
In this section, we first introduce the dataset. Next, we outline a transform for attain-
ing 2D scalograms (images) from 1D ECG signals. Following that, we lead four training
models intended for categorizing scalograms into three categories. Finally, we illustrate
how it works of the suggested technique in Fig. 1.

2.1.1 Dataset
To determine the efficacy of the suggested technique, three separate ECG databases,
ARR, NSR, and CHF, were employed for class labeling. The ARR class has 96 people, the
CHF class has 30, and the NSR class has 36. All raw ECG records have been collected
with a duration of 60 min and a 128-Hz sampling rate [26].

2.1.2 Continuous Wavelet Transform (CWT)

To analyze the signals in both the time and frequency domain, CWT is a great choice
[15]. By using CWT, we can convert 1D signals into 2D scalograms, which give both spa-
tial and temporal information. To acquire the modified signal, CWT convolves it with
the wavelet functions of various scales with translation. CWT of a function f(t) is given
by
 ∞  
t−m
CWT(l, m) = f (t)ψl,m dt. (1)
−∞ l

where ψl,m t−m indicates the shifted and scaled version of the mother wavelet function
 
l
ψl,m (t), and the shifted parameter is m & the scaling parameter is l. Equation (1) gener-
ates a matrix with wavelet coefficients at different scales and locations, and
Figure 2a,b, and c displays the conquered scalograms for three classes via CWT.

2.1.3 Pre‑trained neural networks

DNNs are used for deploying classification models on resource-constrained devices like
mobiles, especially for real-time signal analysis. Popular models, tailored for mobile
compatibility, are evaluated based on accuracy and parameters. Employing transfer
learning minimizes training time, with models initially trained on ImageNet and then
adapted for ECG signal classification. Pre-trained networks used in this paper are
AlexNet, Google Net, Vgg16, and SqueezeNet.

Fig. 2 2D scalograms of: a ARR, b CHF, c NSR

Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 6 of 15

2.1.4 ECG classification
To achieve better classification accuracy, SqueezeNet is used as a pre-trained network and
its block diagram and fire modules are displayed in Figs. 3 and 4. The fire module serves as
the primary element of the network. In this network, the ReLU layer follows each convolu-
tional layer. Conv1 is a convolutional layer with a 3x3 kernel and a stride of 2. For pooling,
3x3 max pooling layers with a stride of 2 are implemented.
SqueezeNet, an early compact model, achieves ImageNet-level accuracy like AlexNet
with 50 times fewer parameters. It emphasizes small kernel sizes (1x1 and 3x3), notably
employing fire modules that compress data via 1x1 convolution before expanding it with
concatenated 1x1 and 3x3 convolutions. Originally comprising 1.2 million parameters
across 68 layers, it is tailored for a 227 x 227 x 3 input image size, classifying into three
categories. Modifications include adjusting dropout layers in the final six layers to 0.6,
resulting in a current model with 724K parameters and 68 layers. The proposed CNN
architecture processes input images through a series of convolutional, activation, pool-
ing, and concatenation layers, culminating in classification outputs.
To compare the performance of the projected CNN algorithm, AlexNet, GoogleNet,
Vgg16, and SqueezeNet are used. The deep learning models with CWT-AlexNet, CWT-
GoogleNet, CWT-Vgg16, and CWT-SqueezeNet are referred to as TM1, TM2, TM3,
and TM4, respectively.

3 Results and discussion

For the performance evaluation, the parameters accuracy (Acc), sensitivity (Se), speci-
ficity (Sp), Precision, F1- score, and Detection Error Rate (DER) are calculated by the
mathematical equations from equation (2) to equation(7).

(TP′1 + TN′1 )
Acc = . (2)
(TP′1 + TN′1 ) + (FP′1 + FN′1 )

(TP′1 )
Se = . (3)
(TP′1 + FN′1 )

(TN′1 )
Sp = . (4)
(FP′1 + TN′1 )

(TP′1 )
Precision = . (5)
(TP′1 + FP′1 )

2 ∗ (TP′1 )
F1 Score = . (6)
2 ∗ (TP1 ) + (FP′1 + FN′1 )
′

(FP′1 + FN′1 )
DER = . (7)
(TP′1 + TN′1 ) + (FP′1 + FN′1 )

where TP′1 = True positive beats FN′1 = False negative beats TN′1 = True negative beats
FP′1 = False positive beats
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 7 of 15

Fig. 3 Architecture of SqueezeNet

Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 8 of 15

Fig. 4 Squeezenet fire module

Table 1 Hyperparameters and their values

Hyperparameters Values

Learning rate 10−4

Mini batch size 20
Max epochs 8
Validation frequency 10
Optimizer Adaptive
moment
estimation

The hyperparameters used in the network are shown in Table 1, and the classification
results in terms of various beats are presented in Table 2. Table 3 describes each class
performance measures such as Acc, Se, Sp, F1-score, Precision, and DER for NSR, ARR,
and CHF, respectively.
The average values of the CWT-SqueezeNet TM4 model yield 98.7% of Acc, 99.1% of
Se, 99.20% of Sp, 98.33% of F1-Score, 98.67% of Precision, and 0.89% of DER, which is
clearly displayed in Table 4.
The proposed method demonstrated outstanding classification results across all
classes and surpassed other methods in detecting arrhythmia. Consequently, the results
accomplished in the current works are analogous to utmost allied works in deep learning
experiments on a similar dataset. Bhati et al. [27] addressed a hybrid DL method com-
bining CNN and BLSTM to categorize the MIT-BIH database, achieving a classification
accuracy of 97.1%. In Kishore et al. [28], a pre-trained EfficientNet B0 CNN architecture
was employed. To obtain 2D scalograms from 1D ECG records, STFT is applied and the
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 9 of 15

Table 2 Simulation results in terms of classification beats

Classifier Classes TP′1 TN′1 FP′1 FN′1

TM1 ARR​ 46 98 4 2
CHF 47 98 3 2
NSR 50 51 0 3
TM2 ARR​ 49 98 1 2
CHF 49 99 3 1
NSR 49 100 0 0
TM3 ARR​ 49 97 1 3
CHF 46 99 4 1
NSR 50 98 0 1
TM4 ARR​ 50 98 0 2
CHF 50 100 0 0
NSR 48 100 2 0

Table 3 Performance metrics of various models for individual classes

Classifier Classes Acc(%) Se(%) Sp(%) F1-score(%) Precision(%) DER(%)

TM1 ARR​ 95.83 96 96.1 93.88 92 4

CHF 95.92 96.67 97 94.94 94 3.33
NSR 94.34 67.33 100 97.08 100 2
TM2 ARR​ 94.2 98 98.98 97.02 98 2.667
CHF 97.9 98.67 96.11 96.07 92 3.33
NSR 98 99.33 100 100 100 0.666
TM3 ARR​ 96 97.33 98.99 96.07 98 2
CHF 98 96.77 99 94.84 98 1.33
NSR 100 99.33 99 99.01 98 0.667
TM4 ARR​ 98.7 98.67 100 99.01 100 1.33
CHF 100 100 100 98.03 100 0
NSR 100 98.67 97.67 97.96 96 1.33

Table 4 Average values of the performance measures for various models

Proposed Acc(%) Se(%) Sp(%) F1-score(%) Precision(%) DER(%)
models

TM1 95.3 86.6 97.7 95.33 95.33 2.44

TM2 96.7 98.7 98.67 97.69 96.67 2.22
TM3 98.0 98.8 98.99 97.99 98 1.33
TM4 98.7 99.1 99.20 98.33 98.67 0.89

achieved classification accuracy is 97.3% when testing the model on the standard Phys-
ioNet database.
Karthiga et al. [29] implemented a deep CNN for ECG signal classification, report-
ing a classification accuracy of 91.92%. Similarly a hybrid CNN-LSTM was intro-
duced [30] for detecting cardiac disorders of ECG records through CWT images.
The authors [31, 32] developed a new classification mechanism with multistage Deep
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 10 of 15

Table 5 Summary of most allied mechanisms and present work

Method Data Input Model Accuracy

Altan et al. [33] 1D ECG DBN 97.3%

Strodthoff et al. [2] 1D ECG ResNet and Inception 89.8%
Siddique et al. [3] 1D ECG Deep extreme machine learning 92.4%
Hsu et al. [5] 1D ECG AlexNet and ResNet 94.4%
Mewada et al. [25] 1D ECG CNN layers 95.2%
Gaddam et al. [17] 2D ECG AlexNet 95.6 %
Panda et al. [22] 1D ECG CNN with EWT 97.6 %
Rahman et al. [20] 1D ECG Alexnet 98.3 %
Ting et al. [13] 2D ECG EfficientNet B0 97.1 %
Bhatia et al. [27] 1D ECG LSTM+CNN 98.3 %
Karthiga et al. [29] 2D ECG CNN 91.9 %
Madan et al. [30] 2D ECG LSTM+CNN 97.3 %
Proposed TM4 2D ECG Squeezenet 98.7 %
Bold values are used to differentiate proposed method results with used signal, model and obtained accuracy

Fig. 5 Performance measures versus training models

Belief Network (DBN), and Altan et al. [33] produced an integrated method for the
early detection of CHF. In this method, Hilbert Huang Transform (HHT) extracts the
features and neural network classifies the data into five classes with an accuracy of
97.53% According to the stated results, the recommended method has 98.7% accuracy
for ARR, 100% for CHF, and 100% for NSR. When compared to prior research that
used other DL models on the same dataset, our model performed better in classify-
ing ECG signals into ARR, CHF, and NSR classes. An overview of the most relevant
studies carried out on the same database is provided in Table 5. The acquired out-
comes are superior to the state-of -the-art methods and will considerably minimize
the amount of contribution required by doctors.
Among all those four models TM4 gives an outstanding performance. Its average
Acc is increased by 3.6%, 2.1%, and 0.7% to TM1, TM2 , and TM3 models. Se value is
enhanced by 14.4%, 0.4%, and 0.3% than TM1,TM2 , and TM3 models, etc. But DER
is reduced by 63.5%, 59.9%, and 3.5% than TM1, TM2 , and TM3 models. Finally, the
entire model’s performance analysis and their graphical representations are shown in
Figs. 5 and 6. The corresponding confusion matrixes and their training progress for
all models TM1, TM2 , TM3 , and TM4 , are shown in Figs. 7, 8, 9, 10, and 11. Confusion
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 11 of 15

Fig. 6 Training models versus detection error rate

Fig. 7 Confusion matrix of TM1 model

matrices are important tools for visualizing and analyzing a model’s predictions
against the actual ground truth across several classes.

4 Conclusion
Arrhythmia classification is an incredibly important topic in healthcare. An arrhyth-
mia is an erratic heartbeat or rhythm. This article presents a deep learning algorithm
that uses alternative time-frequency representations of short ECG segments to reli-
ably diagnose arrhythmia. CWT-RGB using a recommended 2D-CNN model has the
best classification performance concerning average accuracy, sensitivity, specificity,
F1-score, Precision, and DER at 98.7%, 99.1%, 99.2%, 98.3%, 98.7%, and 0.89%, respec-
tively. The confusion matrix of the testing dataset exposed 96.2%, 100%, and 100%
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 12 of 15

Fig. 8 Confusion matrix of TM2 model

Fig. 9 Confusion matrix of TM3 model

validation accuracy for “ARR,” “CHF,” and “NSR.” Furthermore, Se, Sp, F1-score, Pre-
cision, and DER for ARR are 98.67%, 100%, 99.01%, 100%, and 1.33% Se, Sp, F1-score,
Precision, and DER for CHF are 100%, 100%, 98.03%, 100%, and 0%, and for NSR,
Se, Sp, F1-score, Precision, and DER are 98.67% and 97.67%,97.96%,96%, and 1.33%,
respectively. Similarly, the proposed methodology will help doctors accurately iden-
tify cardiac disorders during normal ECG monitoring. Furthermore, the high com-
puting overhead imposed by the usage of CWT is a disadvantage. We could never
attain total intersubjectivity, even if doing so would greatly reduce the scope of action
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 13 of 15

Fig. 10 Confusion matrix of TM4 model

Fig. 11 Training progress of TM4 model

expected by physicians. It will be a fantastic upcoming opportunity for scholars. To

deal with these issues, a strong arrhythmia classification framework is needed.

Author contributions
All authors participated equally.

Funding
Not applicable.
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 14 of 15

Availability of data and materials

Not applicable.

Code availability
Not applicable.

Declarations
Ethics approval and consent to participate
Not applicable.

Consent for publication

Yes.

Competing interests
Not applicable.

Received: 19 August 2024 Accepted: 9 December 2024

References
1. Xu, IoT-assisted ECG monitoring framework with secure data transmission for health care applications, IEEE Access,
vol. 8, Article ID 74586, (2020)
2. N. Strodthoff, P. Wagner, T. Schaeffter, W. Samek, Deep learning for ECG analysis: benchmarks and insights from PTB-
XL. IEEE J Biomed Health Inf. 25(5), 1519–1528 (2021). https://​doi.​org/​10.​1109/​JBHI.​2020.​30229​89. (Epub 2021 May
11 PMID: 32903191)
3. A. Siddiqui, M.A. Athar, S. Khan, Y. Abbas, M.F. Khan, M. Hussain, Modelling, simulation and optimization of diag-
nosis cardiovascular disease using computational intelligence approaches. Jour. Med. Imaging Health Inf. 10(5),
1005–1022 (2020)
4. M. Tantawi, H.A. Shaker, M.F. Tolba. Shedeed, Generalization of convolutional neural networks for ECG classification
using generative adversarial networks. IEEE Access 8, 35592–35605 (2020). https://​doi.​org/​10.​1109/​ACCESS.​2020.​
29747​12
5. P. Hsu, H.L. Liu, Fold electrocardiogram into a fingerprint, in Proceedings of the IEEE/CVF Conf. on Comp. Vision and
Pattern Recog. Workshops, pp. 828-829, Seattle, WA, USA, July, (2020)
6. K. Elleithy, A.-S. Wasimuddin, M. Abuzneid, O. Abuzaghleh. Faezipour, Stages-based ECG signal analysis from tradi-
tional signal processing to machine learning approaches: a survey. IEEE Access 8, 177782–177803 (2020). https://​
doi.​org/​10.​1109/​ACCESS.​2020.​30269​68
7. M. Elgendi, C. Menon, Machine learning ranks ECG as an optimal wearable biosignal for assessing driving stress. IEEE
Access 8, 34362 (2020)
8. X. Gao, Diagnosing Abnormal Electrocardiogram (ECG) via Deep Learning (In Electrocardiography, Intech Open, 2019)
9. Z. Ebrahimi, M. Loni, M. Daneshtalab, A. Gharehbaghi, A review on deep learning methods for ECG arrhythmia clas-
sification. Exp. Sys. App. X 7, 100033 (2020)
10. O. Yildirim, U.B. Baloglu, R.S. Tan, U.R. Acharya, A new approach for arrhythmia classification using deep coded
features and LSTM networks. Comp. Meth. and Prog. Biomed. 176, 121–133 (2019)
11. U.R. Acharya, F. Hamido, L.O. Shu, H. Yuki, H.T. Jen, A. Muhammad, Application of deep CNN for automated detection
of myocardial infarction using ECG signals. Inf. Sci. 415, 190–198 (2017)
12. P.N. Malleswari, C.H. Bindu, K.S. Prasad, Cardiac severity classification using pre-trained neural networks. Interdis. Sci.
Comput. Life Sci. 13, 443–450 (2021)
13. Wei Ting-Ruen, Lu. Senbao, Yuling Yan, Automated atrial fibrillation detection with ECG. Bioengineering 9(10), 523
(2022)
14. A.Y. Hannun, P. Rajpurkar, M. Haghpanahi, G.H. Tison, C. Bourn, M.P. Turakhia, A.Y. Ng, Cardiologist-level arrhythmia
detection and classification in ambulatory electrocardiograms using a deep neural network. Nat. Med. 25(1), 65–69
(2019)
15. U. Amin, S.M. Anwar, M. Bilal, M.M. Raja, Classification of Ar Rhythmia by using deep learning with 2-D ECG spectral
image rep. Remote Sens. 12, 1685 (2020)
16. S. Hong, Y. Zhou, J. Shang, C. Xiao, J. Sun, S. Hu, Opportunities and challenges of deep learning methods for electro-
cardiogram data: a systematic review. Comp. Bio. Med. 122, 103801 (2020)
17. R.V. Sreehari. Gaddam, Automatic classification of cardiac arrhythmias based on ECG signals using transferred deep
learning CNN. J. Phys. Conf. Series 2089(1), 012058 (2021)
18. L. Jia, S. Yujuan, X. Tao, J. Aibiao, Deep convolutional neural network based ECG classification system using informa-
tion fusion and one-hot encoding technique. Math. Prob. Eng. 2018, 7354081 (2018)
19. P.N. Malleswari, C.H. Bindu, K.S. Prasad, A hybrid EMD-DWT based algorithm for detection of QRS complex in
electrocardiogram signal. J. Ambient Intel. Human Comput. 13, 5819–5827 (2022). https://​doi.​org/​10.​1007/​
s12652-​021-​03268-9
20. A.U. Rahman, R.N. Asif, K. Sultan, S.A. Alsaif, S. Abbas, M.A. Khan, A. Mosavi, ECG classification for detecting ECG
arrhythmia empowered with deep learning approaches. Comp. Int. Neurosci. 2022, 6852845 (2022)
21. Y. Liang, S. Yin, Q. Tang, Z. Zheng, M. Elgendi, Z. Chen, Deep learning algorithm classifies heartbeat events based on
ECG signals. Front. Physiol. 11, 569050 (2020)
Malleswari et al. EURASIP Journal on Advances in Signal Processing (2024) 2024:104 Page 15 of 15

22. S. Panda, R.K. Jain, U.R. Acharya. Tripathy, Detection of shockable ventricular cardiac arrhythmias from computational
intelligence and neuroscience 11 ECG signals using FFREWT filter-bank and deep convolutional neural network.
Comp. Bio. Med. 124, 103939 (2020)
23. P. Kanani, M. Padole, ECG heartbeat arrhythmia classification using time series augmented signals and deep learning
approach. Pro. Comp. Sci. 171, 524–531 (2020)
24. M. Thomas, M.K. Das, S. Ari, Automatic ECG arrhythmia classification using DT-CWT based features. Int. J. Electron.
Commun. 69(4), 715–721 (2015)
25. H. Mewada, I.M. Pires, Electrocardiogram signal classification using lightweight DNN for mobile devices. Proc. Comp.
Sci. 224, 558–564 (2023)
26. http://​www.​physi​onet.​org/​physi​obank/​datab​ase/​mitdb: The MIT-BIH Arrhythmia Database
27. S. Bhatia, S.K. Pandey, A. Kumar, A. Alshuhail, Classification of electrocardiogram signals based on hybrid deep learn-
ing models. Sustainability 14(24), 16572 (2022)
28. B. Kishore, A.N.G. Reddy, A.K. Chillara, W.A. Hatamleh, K.D. Haouam, R. Verma, H.K. Atiglah, An innovative machine
learning approach for classifying ECG signals in healthcare devices. J. Healthcare Eng. (2022). https://​doi.​org/​10.​
1155/​2022/​71944​19
29. S. Karthiga, A.M. Abirami, Deep learning convolutional neural network for ECG signal classification aggregated using
IoT. Comp. Syst. Sci. Eng. (2022). https://​doi.​org/​10.​32604/​csse.​2022.​021935
30. P. Madan, V. Singh, D.P. Singh, M. Diwakar, B. Pant, A. Kishor, A hybrid deep learning approach for ECG-based arrhyth-
mia classification. Bioengineering 9(4), 152 (2022)
31. G. Altan, N. Allahverdi, Y. Kutlu, A multistage deep learning algorithm for detecting Arrhythmia. In 1st International
Conference on Computer Applications & Information Security (ICCAIS), pp. 1-5 (2018)
32. N. Allahverdi, Y. Kutlu, G. Altanı, A Multistage Deep Learning Algorithm for Detecting Arrhythmia. IEEE (2017)
33. G. Altan, Y. Kutlu, N. Allahverdı, A multistage deep belief networks application on arrhythmia classification. Int. J.
Intel. Syst. Appl. Eng. 4(Special Issue–1), 222–228 (2016)

Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.