Arrhythmia Classification Techniques Using Deep Neural Network

,


Introduction
Cardiac disorder may cause a severe warning to public health, and in most cases, some arrhythmias can cause severe damage or death. According to the data provided by World Health Organization (WHO) [1], approximately 25.6 million people died due to cardiovascular disease in 2020. Electrocardiogram (ECG) test is used as a diagnostic tool in healthcare institutes. e electrodes attached to the patients' body surface can record the heart's electrical signal over time. Figure 1 shows the ECG leads attached to the human's body surface to get the heart's electrical signals. e healthcare professionals manually diagnose the patient heart condition by interpreting the ECG image. In the advent of technology, several automatic diagnostic tools are developed for arrhythmias classification and detection to assist doctors. PCGs and ECGs are used to diagnose arrhythmia classification. PCG (also known as heart sound auscultation) is commonly listened to or recorded by practitioners through a stethoscope to identify heart irregularities. For this reason, heartbeats have been critically studied to make a diagnosis [2][3][4][5]. Khan et al. [6] provided a very thorough introduction of related studies in the problem domain and proposed a technique based on DNN, for the classification of three kinds of arrhythmia beats. CNN is considered a state-of-the-art tool for the classification of arrhythmia, and it has been studied with several variations such as 1-dimensional, 2-dimensional, or the combination of both [7,8]. According to Xiao et al. [7], a novel arrhythmia classification technique comprises three phases: preprocessing, 1-dimensional CNN architecture based on clique blocks with bidirectional connections between layers and transition blocks with attention mechanism, and majority voting to predict the final result. Experiments were performed on PhysioNet/CinC 2016 database. However, low-frequency noise considerations recorded in arrhythmia beats are ignored, and 2-dimensional acoustical representation requires additional computation, which cannot be accomplished without hyperparameters. Similarly, Noman et al. [8] proposed a framework based on 1-dimensional CNN for direct feature learning from raw arrhythmia beats and 2-dimensional CNN, which takes 2-dimensional timefrequency feature maps.
In the state-of-the-art studies [4][5][6][7][8], automated system which predicts high accuracy results is developed for arrhythmia detection but still not adoptable by healthcare professionals. e primary concerns which affect the success of the developed arrhythmia detection systems are (i) manual features selection, (ii) techniques used for features extraction, and (iii) algorithm used for classification and the most important is the use of imbalanced data for classification. e automated arrhythmia detection required the feature extraction of ECG images that required domain knowledge. In the last few years, deep learningbased systems are being recognized as a tool in healthcare institutes that have capabilities to automatically extract high-level abstract features, avoiding laborious manual feature design. A deep neural network is configured in the same fashion as the human brain works. A single neuron understands and recognizes the pattern based on the logical principles among different components. e proposed work's major concern is to review the recent trends of arrhythmia classification techniques and enlist the limitation and future requirements.
is study can help the researcher get a deeper understanding of arrhythmia classification and the deep learning methods used to develop automated systems. e reminder of the paper is organized into six sections. e methods used in the proposed study are discussed in Section 2, and the research community's related work in the current field is presented in Section 3 with a close comparison. Recent trends of arrhythmia classification and limitation of the current study are discussed in Sections 4 and 5, respectively. e paper is concluded in Section 6.

Methodology
In this research study, the authors follow the mapping techniques for arrhythmia classification techniques using a deep neural network. Figure 2 demonstrates the systematic methodology in a four-step process. e primary objective of this study is to critically review the existing methods based on arrhythmia classification on the publishing forums [9].

Research Objectives.
e primary goal of this research is to review the development of arrhythmias classification techniques over time, i.e., January 2010 to January 2020, using the machine and deep learning approach. e primary objectives of this research study are   e most reliable and trusted data sources available in the digital world to get the required articles for selected research studies are ACM, Elsevier, Google Scholar, IEEE Digital Library, Nature Scientific Reports, Springer, Science Direct, and Web of Science. e two main search strings are used in all data repositories to download the related articles.

Selection Criteria.
Identifying and selecting the bestsuited articles that match the primary objectives of this research study are to identify the latest trends of deep neural networks for arrhythmia classification techniques. Table 1 shows the inclusion and exclusion criteria used in this study.

Literature Review
Based on the selection criteria, the fifty technical articles on arrhythmia classification are examined which are published between January 2010 to January 2020. e selected articles are critically examined, and if any selected articles are available on more than one scientific repository or database, it is considered only once. e different kinds of methodologies, classification algorithms with their accuracy results, and optimization methods used for arrhythmia classification are reviewed from the selected articles. Table 2 shows the literature survey on arrhythmia classification techniques that are used in this study. e authors present the recent trends for arrhythmia classification, the techniques used for features extraction, and the variation of deep neural networks. e study is beneficial for the scientific community to select the arrhythmia classification techniques as per their desires. Table 3 identified the latest trends from recently proposed studies for arrhythmia classification.

Recent Trends in Arrhythmia Classification
e primary objective of this study is to present the techniques used in arrhythmia classification with a publicly available ECG database used in deep/machine learning algorithms. e selected state-of-the-art research studies show the most common methods used for arrhythmia classification and their accuracy results. e accuracy results represent the learning algorithm's success rate that how well the machine was trained to identify the ground truth automatically. e primary objective of this study is to help the scientific community to list the methodologies used in arrhythmia classification with prediction results that researchers can easily select the techniques as per their requirements. Figure 3 shows the statistics of most adopted arrhythmia classification techniques used in selected studies. e deep neural network's fundamental unit is a neuron trained by repetitive tasks and gets experienced just like a human brain through acquired knowledge attained in training. e focus of training and acquiring knowledge is to establish a connection between input and output. After training, the system is capable of   Figure 4 shows the internal working of RNN architecture.

Long Short-Term Memory (LSTM)
. LSTM is the reproduction of RNN architecture [59] commonly used in the field of deep learning. In general feedforward, neurons have forward connections, but LSTM has reversible feedback connections. LSTM processes data points like images or it has the capabilities to process the sequences of data, such as continuous images, video, or speech signals. LSTM architecture has the capabilities to recognize unsegmented speech and detect similar patterns. LSTM also has the ability to recognize the heart sounds signals or ECG beats. e basic architecture of LSTM contains four major elements, i.e., cell, input gate, output gate, and forget gate. e cell has a responsibility to remember the values over specific time intervals and the gates regulate the flow of information between the cells. Figure 5 shows the internal working of LSTM architecture.

Autoencoder.
Autoencoder belongs to the family of artificial neural network-(ANN-) based architecture, which is used for training efficient data coding in an unsupervised manner. ey are recognized as a tool to learn basic patterns for a similar set of data. Autoencoder also performs dimensionality reduction, and it tries to generate a class similar to its original input. In the cited research studies, the author proposed different variations in autoencoder-based architecture. Figure 6 shows the basic architecture of the autoencoder.

Convolutional Neural Network (CNN). CNN is a deep
learning-based algorithm. CNN is widely recognized as a tool in the field of computer vision and image processing. It consists of an input layer, hidden layers, and the last output layer. In forward pass convolutional neural network, the middle layers are called hidden layers that are masked with activation function (ReLU), pooling layer, and convolution. Figure 7 shows the structure of the forward pass convolutional neural network. In the state-of-the-art research studies, CNN is widely recognized as a tool for the detection of ECG beats signals.

Deep Neural Network (DNN).
A simple DNN contains multiple hidden layers that can process the input to output layers [60]. e DNN can recognize different kinds of unstructured data. In arrhythmia classification, authors proposed different kinds of neural networks but the proposed network is composed of the same components: neuron, weight, bias, and function. All these components are capable and act just like the human brain. e deep neural network is widely recognized approach for ECG image classification. Figure 8 shows the basic structure of DNN-based architecture.  connection between all the layers in the network but not with units of each other layers. DBN can be trained by supervision to achieve better prediction. Figure 9 shows the basic structure of DBN-based architecture.

ECG Databases.
e expert systems for automatic detection of arrhythmia disease required training data to understand different patterns of each class. Authors of the selected research study critically analyze arrhythmia classification systems and enlist the most cited/publicly open access ECG databases. e ECG databases available for the scientific community used as a data source are listed in Table 4. e database as mentioned above contains variations of ECG beats with different types of arrhythmia classification. e data sources have unique importance with domain-specific characteristics. e description of the patient population in which these ECG beats were obtained is important in interpreting the methodology and clinical utility in context.

Limitations
Arrhythmia classification techniques are critically analyzed; the researcher puts lots of effort into developing the automated system with good accuracy results.
e developed system may act as clinical support for the healthcare professionals but with having few limitations. e majority of the selected systems are limited to the following: (1) e most ECG databases are not specific to their clinical context. (2) e description of the patient population in which these ECGs were obtained is lacking. is is important in interpreting the methodology and clinical utility in context.     Figure 7: Structure of a forward-pass CNN.

Complexity
(3) e algorithms are trained based on specific environments, and the generalized methodologies are ignored.

Conclusion
Cardiac disorder or arrhythmia is the most dangerous disease that leads to human death. e researcher proposed lots of arrhythmia classification systems to assist the doctors every year. e automated systems predict the high accuracy results used for arrhythmia detection but still not adoptable by healthcare professionals because in the recent studies, authors used time-series data, which is not adaptable in different application environments. Moreover, ECG's timeseries data with signal leads are not appropriate for stable baseline wanders, muscle contraction, and power line interface. In general, the practical methods generally adopted by a cardiologist for screening cardiac patients are 12 leadsbased ECG images. e major concerns that affect the success of the developed arrhythmia detection systems are (i) manual features selection, (ii) techniques used for features extraction, and (iii) algorithm used for classification and the most important is the use of imbalanced data for classification. e automated arrhythmia detection required the feature extraction of ECG images that required domain knowledge. Further, the balanced dataset used for classification methods is required to avoid overfitting. In the last few years, deep learning-based systems are being recognized as a tool in healthcare institutes that have capabilities to automatically extract high-level abstract features, avoiding laborious manual feature design. Finally, this study can help the researcher get a deeper understanding of arrhythmia classification and the deep learning methods used to develop automated systems.
Data Availability e data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.