ACAC: Automatic Cardiac Arrhythmia

Classification Based on 1D-Deep Resnet

M. Dhilsath Fathima S Sithsabesan Jayanthi K
Department of Mathematics Department of Computer Science and
0000-0002-4491-4352 VelTech Multitech Dr. Rangarajan Technology, SRM Institute of Science
Dr.Sagunthala Engineering College and Technology
Dhanyaa S U Chennai, Tamil Nadu, India Tiruchirappalli, Tamil Nadu, India
Department of Computer Science and [email protected] [email protected]
Technology, SRM Institute of Science
and Technology
Tiruchirappalli, Tamil Nadu, India
Chennai, Tamil Nadu, India

Abstract—Arrhythmias are infrequent irregularities in a arrhythmias can lead to heart failure or even mortality [1].
person's heartbeat rhythm. They result from any disruption in Arrhythmias are typically categorized into five groups:
the ordered pattern of the heart's expanding wave of fusion (F), supraventricular ectopic (S), ventricular ectopic
excitation, a process heavily reliant on electrical activity. These (V), non-ectopic (N), and unknown/unclassified (Q) beats
arrhythmias might result in consequences that could be fatal [2]. While some of these categories are non-life-threatening,
and constitute a direct risk to life. As a result, detecting and their unnoticed presence can lead to complications. Thus,
classifying arrhythmias is a significant challenge in cardiac early identification of irregular cardiac rhythms is essential
diagnostics. Electrocardiograms (ECG) are widely used in the
for reducing CVD mortality and providing appropriate
diagnosis of arrhythmias because they are a low-cost, non-
invasive, and rapid technique of diagnosis. However, the
treatments [3]. Electrocardiography (ECG) has become the
unpredictability of arrhythmic segments, along with the most widely used diagnostic tool among medical
sensitivity of ECGs to noise, leads to arrhythmia misdiagnosis. practitioners for monitoring the functioning of the
Furthermore, manually identifying cardiac arrhythmias using cardiovascular system. An ECG signal consists of three
ECG data takes time and is liable for inaccuracy. Deep primary components: the P wave (depicting atrial
learning (DL) is a preferred approach for fast and automatic depolarization), the QRS complex (indicating ventricular
ECG signal classification, surpassing the performance of depolarization), and the T wave (representing ventricular
traditional machine learning models. The suggested study repolarization) [4]. These components convey vital
proposes a novel deep learning architecture, namely 1- information about the heart's electrical conduction system.
dimensional deep residual neural network (1D-deep ResNet) However, due to the nonlinear and noisy nature of the ECG
for an Automatic cardiac arrhythmia classification (ACAC) signal, variations in patterns, and the large amount of data,
system. ECG signals from the MIT-BIH dataset were used to manually classifying heartbeats becomes a challenging and
train and evaluate the model. This proposed model contributes time-consuming endeavor for doctors. The field of
in two different ways. First, the unbalanced ECG data are healthcare has witnessed a significant surge in research,
corrected using the upsampling approach to reduce noise and driven by advancements in machine technology and
prevent biased prediction outcomes. Next, the unbalanced ECG
scientific methods. This growth provides ample
signals are automatically classified using 1D-Deep ResNet.
ACAC system aims to overcome the issues of traditional
opportunities for further development in this domain.
electrocardiograms (ECGs) in the identification of Various Automatic cardiac arrhythmia classification
arrhythmias, which can be impacted by noise and (ACAC) system has been proposed in the literature review.
unpredictability of events, leading to misinterpretation and The ACAC system serves as an essential tool in supporting
mistakes. The confusion matrix is used to calculate the model doctors during the decision-making process. This remains
accuracy, AUC, precision, recall, and f1 score to evaluate the particularly relevant in the domain of automated cardiac
model performance. The results of the experiments shows that arrhythmia classification, where ACAC assists healthcare
the proposed method performed well, with 99.9% and 99.94% professionals in categorizing different types of arrhythmias.
AUC score in the training and testing datasets, respectively. The adoption of ACAC helps address challenges like
The proposed model outperforms other existing deep learning misdiagnosis, human errors, and skill limitations. As a
approaches like CNN, LSTM, and GRU in terms of result, ACAC systems are strongly recommended for
performance, and it will significantly minimise the involvement continuous and prolonged patient monitoring [5]. Machine
of doctors in the classification of ECG signals. According to our learning (ML approaches continually show their efficacy in
research findings, the 1D-deep ResNet is more suitable for improving different ACAC systems [6, 7, 8]. Although these
automated cardiac arrhythmia classification in the future than ACAC models perform predictions, they are not very
other deep learning algorithms now available.
accurate. One of the significant limitations of ML technique
Keywords—1D-deep ResNet, deep neural network, is their manually designed features and shallow feature
Arrhythmia classification, Residual network, cardiac disease. learning architecture. To address this limitation, this
proposed model introduces an innovative deep neural
INTRODUCTION network-based heartbeat classifier called 1D-deep ResNet
In recent years, the rise in cardiovascular diseases capable of both extracting and categorizing heartbeat signals
(CVDs) due to widespread unhealthy lifestyles has become a and predict cardiac problems at an early stage. The main
concerning trend. Among these conditions, the occurrence goal of the 1D-deep ResNet model is to develop the ACAC
of arrhythmia has increased significantly. If left untreated, that analyses samples, categorises distinct types of

979-8-3503-4891-0/23/$31.00 ©2023 IEEE

heartbeats correctly with few or no misclassifications, and arrhythmia classification approach integrating a
achieves high prediction accuracy. metaheuristic optimization (MHO) algorithm with ML
classifiers. It comprises three steps: ECG signal
The proposed model addresses three issues: unbalanced
preprocessing, feature extraction, and classification.
data, low prediction accuracy, and manual feature selection.
Supervised ML classifiers are optimized using the MHO
Imbalanced data distribution is a common problem in
algorithm. A limitation is the need to investigate other
machine learning and deep learning models. It typically
optimization algorithms in future work to enhance heart
occurs when the data in one class is much higher or lower
problem prediction. Real-time monitoring of cardiac patients
than the data in the other classes [9]. To address the issue of
and the incorporation of powerful classification models,
data imbalance, the 1D-deep ResNet classifier utlizes data
such as deep learning, are potential next steps to further
upsampling. Previous ACAC models, based on machine
improve classification accuracy.
learning, used signal processing techniques like Fourier
transform, threshold-based algorithms, wavelet transform, Wang et al. [14] developed a Multi-Head Self-Attention
and Principal Component Analysis (PCA) to extract Mechanism for categorizing arrhythmias from ECG signals.
features. These features were subsequently utilized in ML This model uses a multi-head self-attention mechanism to
models to categorize heartbeats. In traditional machine capture intricate relationships within the data, significantly
learning methods, obtaining relevant features is a manual enhancing classification accuracy. Wavelet transform is
process. Moreover, these techniques demand expert utilized for data preprocessing to reduce noise and enhance
knowledge concerning input variables and specific signal quality, while a linear projection layer extracts
attributes. The involvement of domain experts in feature essential semantic features from the ECG data. This model
extraction is time-consuming, and the resulting features may may demand significant computational resources due to the
be susceptible to noise, translations, and scaling, making multi-head self-attention mechanism, potentially affecting
them less suitable when dealing with new datasets [10]. real-time applications. Prabha et al. [15] proposed an
Deep learning (DL) models have showcased superior arrhythmia prediction model using ECG signals. This model
efficiency in analyzing data and addressing complex utilizes Empirical Mode Decomposition (EMD) for noise
problems by automatically extracting high-level features. removal, Pan Tompkins algorithm for feature extraction, and
This transition has been primarily motivated by the complex Principal Component Analysis (PCA) for signal
nature of ECG signals, an area where conventional machine characteristic reduction. Two classifiers, Support Vector
learning algorithms often face challenges in classification Machine (SVM) and Naive Bayes (NB), are used to identify
tasks [11]. cardiac abnormalities. Comparative analysis shows 89%
accuracy for SVM and 99% for NB. With a substantial
Motivation dataset available, the system is capable of efficient ECG
The proposed ACAC model motivation is to enhance the signal analysis. However, a drawback of this model is its
detection and classification of cardiac arrhythmias. These reliance on traditional machine learning classifiers, which
irregular heart rhythms present significant health risks, may not capture complex patterns and relationships in ECG
underlining the importance of precise identification. While data as effectively as advanced deep learning methods. The
electrocardiograms (ECG) are commonly used for diagnosis, model's performance assessment is dataset-specific and may
they are susceptible to noise and demand time-consuming not generalize well across diverse populations and clinical
manual interpretation. DL provides an efficient and settings. Furthermore, it lacks real-time analysis capabilities,
automated solution that outperforms traditional methods. limiting its utility in continuous monitoring and immediate
The study introduces a 1-dimensional deep residual neural intervention scenarios.
network (1D-deep ResNet) to automate arrhythmia Sraitih et al. [16] developed an inter-patient paradigm
classification. It effectively addresses data imbalances and designed to enhance the detection of minority arrhythmia
noise, achieving outstanding results with AUC scores of classes. The researchers evaluated supervised ML models,
99.9% in training and 99.94% in testing. This research aims including SVM, k-Nearest Neighbours (KNN), and Random
to enhance both the efficiency and accuracy of arrhythmia Forest (RF), using ECG data from MIT-DB. The evaluation
classification, potentially reducing the need for manual criteria encompassed essential metrics such as accuracy,
intervention and improving the quality of patient care. precision, recall, and f1-score. Notably, the SVM classifier
demonstrated superior performance with an accuracy of
0.83. However, it's worth noting that this study
EXISTING DEEP LEARNING TECHNIQUES IN CARDIAC predominantly focuses on the SVM classifier, warranting
ARRHYTHMIA CLASSIFICATION further exploration of the other three methods. Moreover,
This section discusses the development of cardiac the generalizability of these findings to larger and more
arrhythmia classification models based on ML and DL. diverse datasets or their applicability to various clinical
Madan et .al [12] developed hybrid deep learning model for scenarios may require additional investigation.
automating the detection and classification of cardiac Zhang et al. [17] developed an approach to heartbeat
arrhythmias. This method transforms 1D ECG signals into classification that combines hybrid time-frequency analysis
2D Scalogram images, simplifying noise filtering and and ResNet-based transfer learning. This method involves
feature extraction. The approach combines a 2D
transforming raw ECG signals into spectrogram images
convolutional neural network (CNN) and a Long Short-
using a hybrid time-frequency analysis technique, which
Term Memory (LSTM) network, resulting in a 2D-CNN-
LSTM hybrid model and achieving 98.7% accuracy for captures both temporal and frequency domain information
Cardiac Arrhythmias. A major limitation is the heavy required to distinguish heartbeats. These images are then
computing burden due to the usage of Continuous Wavelet fed into a pre-trained ResNet architecture on large-scale
Transform (CWT). These computational demands can image datasets, which uses transfer learning to extract high-
potentially limit their real-time application in clinical level features relevant to heartbeat classification. While the
settings. Hassaballah et al. [13] developed an automated approach achieves promising results in accurately
distinguishing various types of heartbeats, it is Dataset description
computationally complex, especially during the training The PhysioNet MIT-BIH Arrhythmia dataset [18] are
phase, due to the deep learning architecture and the need for extensively utilised in the field of ECG signal analysis and
extensive ECG signal preprocessing. Furthermore, relying have been used for algorithm development, testing, and
on transfer learning may limit the model's adaptability to comparison in several prior publications. Researchers and
domain-specific features in ECG data, potentially resulting developers can access these datasets as a free resource
in suboptimal performance in scenarios with limited labelled because they are publicly available. Furthermore, as these
data or different data distributions. annotations were decided by several subject-matter experts,
they are regarded as accurate and dependable when included
with these datasets. It is a pre-processed and beat-segmented
Research contribution of the proposed model database that has 109,446 ECG samples with 187 features.
The ACAC model utilizes a 1D-Deep ResNet model, This dataset has 5 classes: 0 = N (Normal Beat) ,1 = S
representing a novel approach to ECG signal (Supraventricular premature beat), 2 = V (Premature
classification. This model combines 1D ventricular contraction), 3 = F (Fusion of ventricular and
convolutional layers with deep residual neural normal beat), 4 = Q (Unclassifiable beat).
networks, enabling it to capture complex features in
ECG signals. This contribution enhances the
accuracy and effectiveness of arrhythmia Fig. 1. The proposed Automatic cardiac arrhythmia classification
classification. (ACAC) system

Class imbalance can lead to biased classification results. Exploratory data analysis (EDA)
The essential step taken by the ACAC model to The EDA approach was performed for identifying an
address this issue is using data resampling on the essential characteristic of the input dataset. The EDA
MIT-BIH dataset. This contribution ensures that the process on the MIT-BIH ECG Arrhythmia dataset, showed
model is trained on a more balanced dataset, which is an imbalanced distribution of class labels. This means that
crucial for reliable arrhythmia classification the number of instances in each class is not evenly
distributed, which can potentially lead to biased prediction
The evaluation of the 1D-Deep ResNet model against
and affect the performance of a classification model.
other deep learning models such as CNN, LSTM, and
Addressing this class imbalance issue is an important step in
GRU demonstrates issts effectiveness. This
developing a robust and accurate ECG arrhythmia
comparison provides insights into the relative
classification model. Table 1 represents the ECG sample
performance of the proposed model, highlighting its
count for each class label.
strengths and advantages over alternative approaches.
ACAC makes use of deep learning methods well- Resampling
recognized for their ability to manage complex data Resampling is a method for handling extremely
patterns, potentially resulting in more precise imbalanced datasets. It involves either adding more
arrhythmia classification. instances from the minority class (over-sampling) or
deleting samples from the majority class (under-sampling).
The model is designed to provide efficient and timely Initially, five different dataframes are created based on the
responses in clinical scenarios, enabling quick values of the target class variable. Subsequently, the major
interventions when required. dataframe is downsampled to 20,000 to balance the
distribution of the target variable. Next, each of the four
dataframes representing the minority classes is upsampled to
OUTLINE OF THE PROPOSED MODEL 20,000 samples. Thus, upsampled minority classes and the
The proposed automated cardiac arrhythmia downsampled majority class are concatenated into a new
classification model (ACAC) system has been developed dataframe for training. Consequently, the distribution of the
using the 1D-deep ResNet model, which combines 1D target variable across all class labels is balanced. Figure 2
convolutional layers with a deep residual neural network for displays the output of resampling, where class label samples
ECG signal classification. The 1D-Deep ResNet model is are equally distributed.
trained and validated using the MIT-BIH dataset. The
ECG SAMPLE COUNT FOR EACH CLASS LABEL
development of the ACAC model consists of two main
stages: In the first stage, data resampling is applied to the Name of
Class Labels Count
MIT-BIH dataset to balance class labels, thereby avoiding Attribute
biased classification. In the second step, the proposed 1D-
Deep ResNet classifier trains and validates the ACAC model 0 72471 Normal Beat
using the balanced data samples from MIT-BIH. The results
obtained from the 1D-Deep ResNet model are evaluated 1 2223 Supraventricular premature beat
alongside other deep learning (DL) models such as CNN,
LSTM, and GRU. To measure the efficiency of the 1D-Deep 2 5788 Premature ventricular contraction
ResNet, various classification performance metrics were Fusion of ventricular and normal
used, including Classifier Accuracy, Misclassification rate, 3 641
beat
Sensitivity, Specificity, Precision, F-Score, and the Receiver
Operating Characteristic curve. The process flow diagram of 4 6431 Unclassifiable beat
this proposed system is depicted in Figure 1.
 Steps to configure 1D-Deep ResNet
Step 1: Dataset Setup
Training Dataset: Denoted as , it consists of

pairs , where is an input feature, and is

the corresponding output class label.
Testing Dataset: Denoted as , similar to the
training dataset, it consists of pairs

Step 2: Skip Connection in ResNet

Fig. 2. MIT-BIH Samples After Equal Distribution using Resampling
In the ResNet architecture, a skip connection can be
mathematically expressed as follows:
1-D Deep Residual Neural Network (1-D Deep
ResNet) represents the input to the next block (Identity or
The 1D-Deep ResNet is a deep neural network that uses Convolutional). is the output of the previous block. is
mathematical operations to convert input states into outputs the input to the previous block.
by computing the probability of each output unitwise. Each
of these mathematical procedures corresponds to one of the Step 3: Cost Function
neural network's layers. A deep stack of calculations is The cost function of the ResNet model aims to minimize the
necessary in complicated builds, thus the name 'Deep variance between the actual and predicted values. The cost
ResNet.' Given that ECG classification characteristics function is represented as follows:
frequently show in the morphology of ECG waveforms,
feature extraction methods must be tuned to characterise
these waveforms using a set of parameters. The name 1-D is the model estimator or the cost function.
Deep ResNet derives from the convolution process in a is an underlying loss function that quantifies
CNN model, which has demonstrated efficacy in feature the cost of predicting when the actual
extraction from 1-D data while providing resilience against
label is
noise. The 1-D Deep ResNet incorporates shortcut links
between its layers to overcome a recognised limitation of Step 4: Evaluation and Interpretation
traditional deep neural networks. The 1-D Deep ResNet
exhibits potential in feature extraction from unprocessed Evaluate the Arrhythmia classification model
ECG waveforms. The two main structural parts of the 1D- using the testing dataset . Interpret the model's
Deep ResNet are the feature learning and classification units performance using various classification performance
[19]. Figure 3 presents the construction of the 1-D Deep measures.
ResNet in detail.
Hyperparameters of 1D-Deep ResNet
The feature learning component is composed of several In deep neural networks, optimizing model
convolutional blocks and identity blocks. The identity block hyperparameters is crucial for achieving high-performance
is used when the input and output dimensions of a block are machine learning models. Hyperparameters are the settings
the same. Its primary function is to maintain the dimensions
and configurations of the model that are not learned from the
and propagate the input features unchanged through the
data but must be specified during model development.
block. This is crucial for preserving important features in the
network. In contrast, the convolutional block is used when Tuning hyperparameters is an essential step in enhancing
the dimensions of the input need to change. It introduces 1D-Deep ResNet algorithms. Table 2 represents
non-linearity and allows the network to learn new features hyperparameters of the proposed 1D-Deep ResNet.
from the input data.
HYPERPARAMETERS OF 1D-DEEP RESNET
The 'filters' parameter specifies the number of filters in
each layer within each block, and each block consists of Classifier Parameter Value
three convolutional layers. The first layer has a kernel size
3
of one, while the second and third layers have kernel sizes of Epochs
three, which aids in feature extraction from the input data.
256
The initial data processing begins with a 1D convolutional Batch size
layer featuring 64 filters and a kernel size of 7. Following
this, a max-pooling layer with a pool size of 3 and a stride of 256
Neurons in input layer
2 is applied. The input then passes through a sequence of 1D-Deep
ResNet blocks, each belonging to different stages. The ResNet 1, 3
Kernal size
number of filters in each block increases as the stage number
advances, allowing the network to capture increasingly Activation functions of Relu
complex features from the data. The network's final output hidden layers
is then processed through global average pooling, which Activation functions of Softmax
aggregates the features, followed by a fully connected layer output layer
with a softmax activation function. This combination results
in the final classification output.
 EXPERIMENTAL RESULTS AND DISCUSSIONS
The balanced input dataset is split two ways, with 80% of
the data to training the 1D-Deep ResNet model and the
remaining 20% used for testing and comparative analysis of
the final Arrhythmia Classification and Analysis
Convolutional model. To avoid overfitting, a 10-fold
repeated and stratified cross-validation process is used. This
proposed model uses a variety of classifier performance
metrics to thoroughly assess the efficacy of the 1D-Deep
ResNet-based cardiac arrhythmia classification model.
This comprehensive evaluation compares the performance Fig 4. Accuracy Graph
of other deep learning models, including Convolutional
Neural Networks (CNN), Long Short-Term Memory
(LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM
(BiLSTM), and traditional machine learning models such as
logistic regression, decision tree, K-nearest neighbours, and
support vector machine. The findings of this evaluation are
shown in Table 3, which displays the respective
performance metrics of each model.
The training the 1D-Deep ResNet for only three epochs with
a batch size of 256 neurons in the input layer and kernel
Fig 5. Loss Graph
sizes of one and three, using ReLU activation functions for
hidden layers and Softmax for the output layer, results in an
outstanding test accuracy of 99.94%. This strong CONCLUSION AND FUTURE PERSPECTIVE
performance demonstrates the effectiveness of the model's
architecture and training parameters. Furthermore, as The ACAC model has demonstrated remarkable
detailed in the data, this performance significantly
performance using the proposed 1D-Deep ResNet.
outperforms that of other models, including BiLSTM, CNN,
Through the integration of 1D convolutional layers with
and other conventional machine learning algorithms such as
Logistic Regression, Decision Tree, KNN, and SVM. deep residual neural networks, 1D-Deep ResNet
classifier has achieved high accuracy and efficiency of
99.94%, even without extensive feature engineering.
1D-Deep ResNet outperforms other model due to its
deep architecture with skip connections, which
facilitates learning hierarchical features efficiently.
Fig. 3. Architecture of the proposed 1D-Deep ResNet ResNet's ability to capture both spatial and temporal
features makes it well-suited for tasks like arrhythmia
COMPARATIVE ANALYSIS OF THE PROPOSED 1D-DEEP classification, especially when datasets contain complex
RESNET CLASSIFIER
patterns or require extracting intricate spatial features.
Training Testing Traini The use of data resampling to address class imbalance
Testing
Model Accuracy Accuracy ng
(%) (%) Loss
Loss has significantly improved the reliability of arrhythmia
Logistic classification. The validation results are incorporated
91 87 0.35 0.46
Regression into the model development pipeline to assess
Decision Tree 95 89 0.22 0.28 performance and prevent overfitting. The comparative
KNN 92 88 0.32 0.29 analysis with other deep learning models, such as CNN,
SVM 71 70 0.76 0.74 LSTM, and GRU, along with traditional machine
BiLSTM 99.2 98.91 0.06 0.07 learning models, has shown the superior performance of
CNN 98.38 95.50 0.04 0.14 the 1D-Deep ResNet.
GRU 71.72. 79.71 0.62 0.70 The hyperparameters of the proposed 1D-Deep ResNet
LSTM 49.48 50.38 1.21 1.19 have been chosen based on several considerations to
1D-Deep
ResNet
99.90 99.94 0.05 0.04 optimize the model's performance. Firstly, setting the
number of epochs to 3 ensures that the model has
sufficient training iterations to learn from the data
without risking overfitting, striking a balance between
Figures 4 and 5 depict the proposed 1D-Deep ResNet model
convergence and computational efficiency. A batch size
accuracy and loss graph, respectively.
of 256 has been selected to balance the trade-off
between computational resources and model
convergence. A larger batch size can accelerate training
due to increased parallelization but may also increase
memory consumption. The choice of 256 neurons in the
input layer is likely determined by the complexity of the
input data and the desired capacity of the model to
 capture relevant features. Using kernel sizes of 1 and 3 Fathima, M. Dhilsath, S. Justin Samuel, and S. P. Raja. "HDDSS: An
Enhanced Heart Disease Decision Support System Using RFE-ABGNB
enables the model to capture both local and global Algorithm." International Journal of Interactive Multimedia & Artificial
features in the data. A kernel size of 1 focuses on Intelligence 527 (2023).
capturing local features, while a kernel size of 3 Asif, Rizwana Naz, Sagheer Abbas, Muhammad Adnan Khan, Kiran
captures broader patterns in the data. Sultan, Maqsood Mahmud, and Amir Mosavi. "Development and validation
of embedded device for electrocardiogram arrhythmia empowered with
The utilization of classifier performance metrics has transfer learning." Computational Intelligence and Neuroscience 2022
provided valuable insights into the model's strengths. (2022).
The outcomes of the 1D-Deep ResNet bring up Dablain, Damien, Bartosz Krawczyk, and Nitesh V. Chawla.
compelling possibilities for study and progress in the "DeepSMOTE: Fusing deep learning and SMOTE for imbalanced
data." IEEE Transactions on Neural Networks and Learning
field of ECG signal categorization. A few suggestions Systems (2022).
for exploration include continuously refining and Asghar, Ali, Saad Jawaid Khan, Fahad Azim, Choudhary Sobhan Shakeel,
extending the ResNet architecture to extract more Amatullah Hussain, and Imran Khan Niazi. "Review on electromyography
intricate features from ECG signals, testing the model's based intention for upper limb control using pattern recognition for human-
machine interaction." Proceedings of the Institution of Mechanical
performance on larger and more diverse datasets to Engineers, Part H: Journal of Engineering in Medicine 236, no. 5 (2022):
ensure its applicability in various clinical scenarios, and 628-645.
investigating the combination of ECG data with other Hu, Rui, Jie Chen, and Li Zhou. "A transformer-based deep neural network
medical data sources, such as imaging or patient for arrhythmia detection using continuous ECG signals." Computers in
Biology and Medicine 144 (2022): 105325.
records, for comprehensive diagnostics.
Madan, Parul, Vijay Singh, Devesh Pratap Singh, Manoj Diwakar, Bhaskar
Pant, and Avadh Kishor. "A hybrid deep learning approach for ECG-based
REFERENCES arrhythmia classification." Bioengineering 9, no. 4 (2022): 152.
Deng, Penghong, Yu Fu, Mingsheng Chen, Dong Wang, and Lei Si. Hassaballah, Mahmoud, Yaser M. Wazery, Ibrahim E. Ibrahim, and Aly
"Temporal trends in inequalities of the burden of cardiovascular disease Farag. "Ecg heartbeat classification using machine learning and
across 186 countries and territories." International Journal for Equity in metaheuristic optimization for smart healthcare
Health 22, no. 1 (2023): 1-13. systems." Bioengineering 10, no. 4 (2023): 429.
Sindhu, Velagapudi Swapna, Kavuri Jaya Lakshmi, Ameya Sanjanita Wang, Yue, Guanci Yang, Shaobo Li, Yang Li, Ling He, and Dan Liu.
Tangellamudi, and K. Ghousiya Begum. "A novel deep neural network "Arrhythmia classification algorithm based on multi-head self-attention
heartbeats classifier for heart health monitoring." International Journal of mechanism." Biomedical Signal Processing and Control 79 (2023): 104206.
Intelligent Networks 4 (2023): 1-10.
Praba, R. Anandha, L. Suganthi, ES Selva Priya, and J. Jeslin Libisha.
Gupta, Kanishk, Santhosh Kumar, Meena Anand Kukkamalla, Vani Taneja, "Efficient Cardiac Arrhythmia Detection Using Machine Learning
Gufaran Ali Syed, Fawaz Pullishery, Mohammad A. Zarbah, Saeed M. Algorithms." In Journal of Physics: Conference Series, vol. 2318, no. 1, p.
Alqahtani, Mohammed A. Alobaoid, and Saurabh Chaturvedi. "Dental 012011. IOP Publishing, 2022.
Management Considerations for Patients with Cardiovascular Disease—A
Sraitih, Mohamed, Younes Jabrane, and Amir Hajjam El Hassani. "An
Narrative Review." Reviews in Cardiovascular Medicine 23, no. 8 (2022):
automated system for ECG arrhythmia detection using machine learning
261.
techniques." Journal of Clinical Medicine 10, no. 22 (2021): 5450.
Itchhaporia, Dipti. "Artificial intelligence in cardiology." Trends in
Zhang, Yatao, Junyan Li, Shoushui Wei, Fengyu Zhou, and Dong Li.
cardiovascular medicine 32, no. 1 (2022): 34-41.
"Heartbeats classification using hybrid time-frequency analysis and transfer
Chen, Chun-Yen, Yan-Ting Lin, Shie-Jue Lee, Wei-Chung Tsai, Tien-Chi learning based on ResNet." IEEE Journal of Biomedical and Health
Huang, Yi-Hsueh Liu, Mu-Chun Cheng, and Chia-Yen Dai. "Automated Informatics 25, no. 11 (2021): 4175-4184.
ECG classification based on 1D deep learning network." Methods 202
MIT-BIH Arrhythmia Database. Available
(2022): 127-135.
online: https://www.physionet.org/content/mitdb/1.0.0/ (accessed on 24
Bharti, Rohit, Aditya Khamparia, Mohammad Shabaz, Gaurav Dhiman, February 2023)
Sagar Pande, and Parneet Singh. "Prediction of heart disease using a
Automatic Cardiac Arrhythmia Classification Using Combination of Deep
combination of machine learning and deep learning." Computational
Residual Network and Bidirectional LSTM" by Runnan He et.al. in IEEE
intelligence and neuroscience 2021 (2021).
Access, doi: 10.1109/ACCESS.2019.2931500)