HybDeepNet: A Hybrid Deep Learning Model for Detecting Cardiac Arrhythmia from ECG Signals

The problem to be addressed is the high mortality rate of heart disease and the need for reliable and early detection techniques to prevent fatalities. Several clinical tests, including electrocardiogram (ECG) signals, heart sound signals, impedance cardiography (ICG), magnetic resonance imaging, and computer tomography can be used to determine whether an individual has heart disease. In this research, three deep learning models - Multi-layer Perceptrons (MLPs), Deep Belief Networks (DBNs)


Introduction
Cardiovascular disease (CVD) is a widespread non-communicable illness that is a major cause of death worldwide. It can result from any anomaly in the cardiovascular system and can manifest in a number of forms such as heart failure, heart attacks, traumatic strokes, infarction strokes, and other arrhythmia and heart muscle-related disorders [4]. According to the World Health Organization, an estimated 16.9 million people die from CVD every year, and it accounts for 29% of all deaths worldwide. Developing nations are particularly affected, with 78% of deaths from CVD occurring in these countries. It is projected that by 2035, the number of annual deaths from CVD will reach 24 million [1]. The main contributing factors are hypertension, an irregular heartbeat, hyperlipidemia, hypoglycemia, and elevated blood pressure. The linked risk factors that may cause CVD are psychological trauma, lack of exercise, tobacco usage, alcoholism, overweight, genetic factors, poor diets, and a lack of physical workouts. The linked risk factors that may cause CVD are psychological trauma, lack of exercise, tobacco usage, alcoholism, overweight, genetic factors, and poor diet [12].
Coronary artery disease (CAD) is one of the most frequent kinds of CVD. This happens due to the abnormality in minimum one of the left anterior dorsal, left circumflex, and right arteries [28]. In CAD, interfacial tissue in the epithelium of the coronary arterial wall mixes with low-density lipoprotein, exposing them to additional lipid modification and aggravation, which causes poor vascular lesions. As irritation progresses, cell interfacial fatty accumulation induces apoptosis in the vascular system, causing high magnesium deposition and ulceration. The carotid arteries' thrombus expands, leading to myocardial luminal blockage, which limits blood circulation and provides the heart with enough blood and oxygen, thus leading to hypoxia in the muscles [13].
Despite having stiffening and fibro muscular coverings, atherogenesis nodules with significantly tiny hydrophobic cores can slowly produce infarction due to progressive atherosclerotic volume increase that infringes the cardiac channel radius [14]. On the other hand, specific liposome arterial plaques have wider cores and weaker fibrous coverings that are prone to collapse, in which the fluids rapidly leak into the myo-cardial stream, causing the emboli to develop. This development could obstruct the inflow of cerebral hypoxia in the cardiomyopathy resulting in Thromboembolic lesions. In this condition, an individual's cardiac muscles perish due to oxygen deprivation for longer [6].
Myocardial infarction is a type of CVD caused due to hypo-perfusion, which harms the heart muscles. Chronic recurrent infarction might result in deleterious reconfiguration of the right atrium and diminished heart muscle contraction [26]. Additionally, physiological consequences of MI, such as rheumatic heart disease, decrease efficiency or breach the posterior segment and atrial septal region, aggravating cardio humiliation and resulting in heart attacks. Timely diagnosis of myocardial infarctions is crucial for the early intervention and prevention of the potential onset of cardiovascular diseases [23].
Cardiovascular disease (CVD) diagnosis involves diagnostic procedures like coronary angiography and screening tests. Additionally, alternative passive cardiac testing methods are also available that have other limitations ranging from uncertainty in detecting appropriate order, duration of the results, and selection of proper cardiac imaging procedures [8]. Further, additionally, other assessments include cardiac neuroimaging or echocardiogram. The cost of echocardiograms and the need for qualified specialists to examine the neuroimaging results are among the limitations of these diagnostic tests. Machine learning techniques have recently been applied more successfully to classify CVDs.
The most prevalent type of CVDs is cardiac arrhythmia, which can be correctly identified from electrocardiogram (ECG) records. The ECG signal detects the irregularity of the heart to identify cardiac arrhythmia. An essential medical tool for automatically detecting CVD is the ECG, which documents the features of cardiac contraction, healing, and susceptibility. Finding abnormal cardiac rhythms in the ECG readings is crucial. Manual examination, which is laborious and time-consuming, is necessary to interpret the ECG recording. As a result, many Machine learning techniques have been used to identify cardiac arrhythmias from ECG signals automatically. The main process control methods needed by convention- al ML systems are feature engineering, pattern discovery, feature representation, and classification [29]. The main limitation of such systems is the selection of accurate features using the correct elements from the ECG signals to detect CVD. In recent times, applications requiring prediction and classification tasks have greatly benefited from the use of Deep Learning approaches, as these models do not suffer from the risk of selecting and extracting features [18]. However, because of the ECG signal's weak amplitude, doctors frequently miss its abnormalities. Therefore, creating trustworthy DL-based models for CAD early detection and robust classification is a difficult challenge. On the identification and categorization of CVD, numerous investigations have been conducted and reported in the literature [10].
Six different pieces of information collected from the patients are used to identify cardiac arrhythmia [25]. These diagnostic tests' characteristics include Electrocardiogram, heartbeat sound, Flow cytometry, MRI, and CT. These six different patient signals can be used for diagnostic purposes. However, each method has advantages and disadvantages for diagnosing cardiac arrhythmia [21]. The main emphasis of this paper is the development of suitable CAD detection models using the participants' ECG signals as input. In light of this, a review of the cardiovascular disease literature has been conducted with a focus on the material relating to ECG-based Cardiac arrhythmia identification [22]. To have a better understanding of the current stream of research on Cardiac arrhythmia, a small number of pertinent recent findings employing different kinds of signals and features are also examined [9].

Contributions to the Existing Work
The main contributions of this work are as follows:

Organization of the Paper
The remainder of the paper is organized as follows. Section 2 reviews the existing works related to cardiovascular disease prediction. Section 3 outlines the proposed hybrid deep neural network model. Section 4 discusses the results of performing experiments using the datasets. Section 5 concludes the current research work

Related Works
The state-of-the-art studies in the existing literature for cardiovascular disease prediction are discussed briefly in this section. Several articles published between 2014 and 2022 are chosen to review the works carried out by several authors for the problem taken. It was discovered that most of the results used signals such as heart sound, Radiography, and Optical coherence tomography. Research on the papers gathered from various sources shows that different forms of pulses from the human body have been used as input to perform cardiovascular disease classification methods.
Sinus arrhythmia was directly determined in [27] using a Recurrent neural network. Accuracy of 97.6% and 98.55% are claimed to be obtained for 30 participants using 10-fold cross-validation and for ten participants using blindfold validation. The deep learning technique based on bi-LSTM is utilized in [30] to categorize the ECG signal. The median filter analysis in the study is used to retrieve the secondary signal's bands, followed by using it in the deep neural networks as input. A 98.54% recognition rate has been achieved for the suggested method. Using coupled CNN and LSTM models, the ECG signal has been detected in [31]. In the initial layers, the CNN is utilized to extract features, and the LSTM learning algorithm is then further employed for classification using the features that the CNN has removed. According to reports, the combined technique offers a better accuracy rate [11]. In [16], four different characteristics are extracted from the ECG signal before input into the LSTM model for diagnosis. The experiments conducted through simulation produce a reasonable level of precision, accuracy, and F1-score. The authors in [25] describe the development of an attributes-based bi-LSTM model for classifying arrhythmia from ECG data. The work in [21] reports on the localization of cardiac rhythm using a temporal LSTM network. It is demonstrated that the suggested paradigm offers convincing performance. Using GAN, suppression of noises in CT has been accomplished in another intriguing research [25]. The simulation study used unbalanced CT data, and the GAN approach's successful application in reducing the noise component was demonstrated. In [17], a mixed hybrid model is proposed, exhibiting an enhanced performance and noise reduction in CT images. The CVD categorization based on Convolutional Neural Networks is claimed to produce an altered ECG signal before feeding it to the CNN model. Moreover, it incorporates Short Time Fourier Transform for preprocessing [3,20].
Support vector machine (SVM) was used by the authors of [32] as a classifier to identify CAD. The optimization algorithms such as genetic and binary PSO have been used to choose the features. The suggested model is trained and tested using the 10-fold validation. It is shown that the proposed strategy performs Not suitable for datasets with more classes better prediction performance while requiring less complexity than the SVM model used without optimization [26]. Table 1 provides a comparative analysis of the existing methods in the literature for cardiovascular disease diagnosis.

Research Gap and Motivation of Research
From the perspective of health care, reliable, precise, and quick detection of cardiovascular diseases is the highest priority. Additionally, it is true that the Deep learning-based detection of Heart diseases generally provides more excellent performance in contrast to the Machine learning techniques. Further, it is observed that in most existing works for heart disease prediction, only one dataset has been utilized to evaluate the effectiveness of the various models for classification. The conclusion drawn from one dataset's findings may not apply to other datasets. Hence, there is necessary to test and assess the performance of the proposed models with at least two standard datasets.
Additionally, it has been noted that the ensemble rendition of the best-performing models shows improved detection accuracy compared to the models employed individually. Several DL models, including MLP, RBM, and DBN, have only been utilized in a few studies. Hence, these models are combined in the present research to detect cardiac arrhythmia from ECG signals

Proposed Methodology
This section outlines the proposed methodology, which involves three deep learning models Multilayer Perceptrons, Restricted Boltzmann Machines, and Deep Belief Networks. These models are mainly developed for this research as these models have been rare for Cardiac arrhythmia prediction using deep learning. Moreover, to achieve high performance, instead of using these models independently, a hybrid approach is proposed to combine these models. MLPs are used to optimize the proposed model's hyperparameters, RBMs extract insightful features from the data, and further, the classification of whether the patient is affected by cardiac arrhythmia or not is performed using the DBNs. The architecture of the proposed hybrid model is presented in Figure 1. to the served iction, te the cation. s may sary to odels semble shows odels luding d in a in the from hyperparameters, RBMs extract insightful features from the data, and further, the classification of whether the patient is affected by cardiac arrhythmia or not is performed using the DBNs. The architecture of the proposed hybrid model is presented in Figure 1.

Figure 1
Proposed methodology

Multilayer Perceptrons (MLPs)
The multilayer perceptron algorithm employs artificial neurons in several layers, including hidden layers. For problems requiring binary classification, these algorithms are used. Algorithms called multilayer perceptrons are derived from the concept of biological neurons. The structure of a Multilayer perceptron is depicted in Figure 2. They employ perceptrons, which are synthetic neurons. Each neuron in a perceptron has an activation function. Each neuron's weighted inputs are mapped by the activation function, which also decreases the number of layers to only two. By changing the weights that are given to it, a perceptron evolves. The parameters that are set for the execution of the Multilayer perceptron algorithm are shown in Table 2.

Multilayer Perceptrons (MLPs)
The multilayer perceptron algorithm employs artificial neurons in several layers, including hidden layers. For problems requiring binary classification, these algorithms are used. Algorithms called multilayer perceptrons are derived from the concept of biological neurons. The structure of a Multilayer perceptron is depicted in Figure 2. They employ perceptrons, which are synthetic neurons. Each neuron in a perceptron has an activation function. Each neuron's weighted inputs are mapped by the activation function, which also decreases the number of layers to only two. By changing the weights that are given to it, a perceptron evolves. The parameters that are set for the execution of the Multilayer perceptron algorithm are shown in Table 2. The algorithm for the Multilayer perceptron is given in Algorithm1.

Multilayer Perceptron Algorithm
Input: Nodes 1 2 , , ., n x x x … in the input layer Output: Nodes , , ., in the output layer Step 1: Initialize b as bias and w as weights Step 2: For each weight w xy Step 3: Step 4: For each bias y σ Step 5: Step 6: For each training sample s in N Step 7: For each unit x in the input layer Step 8: Out Step 9: For each unit y in the hidden layer Step 10: Hid W In b y xy x y = Σ + Step 11: Step 12: For each unit z in the output layer Step 13: Step 14: For each unit y in the hidden layer, moving from the final to initial layer in the hidden unit Step 15:

Restricted Boltzmann Machines (RBMs)
The Boltzmann Machine is a random probabilistic model stacked in two layers. This model's two layers are hidden and viewable in nature. The viewable layer includes a set of states. Similarly, the hidden layer is represented as, The states that relate these two layers are completely packed and interconnected. The structure of the RBM model is presented in Figure  3. The conditions are devoid of connections within themselves. This model is also known as a model based on energy. There exists a multiplicative type of energy between the layers. The model learns by altering this energy variant according to the requirement. The distribution of the probability-based function for the RBM model is denoted as in (1): Structure of Multilayer Perceptron Step 13: Step 14: For each unit y in the hidden layer, moving from the final to initial layer in the hidden unit Step 15:

Figure 2
Structure of Multilayer Perceptron x n denotes the bias variables. The bias variable may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.

Structure of Restricted Boltzmann Machine
The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: (3)

Algorithm 1
Multilayer Perceptron Algorithm Input: Nodes 1 2 , , ., n x x x … in the input layer Output: Nodes , , ., in the output layer Step 1: Initialize b as bias and w as weights Step 2: For each weight w xy Step 3: Step 4: For each bias y σ Step 5: Step 6: For each training sample s in N Step 7: For each unit x in the input layer Step 8: Out In Step 9: For each unit y in the hidden layer Step 10: Hid W In b y xy x y = Σ + Step 11: Step 12: For each unit z in the output layer Step 13: Step 14: For each unit y in the hidden layer, moving from the final to initial layer in the hidden unit Step 15:

Figure 2
Structure of Multilayer Perceptron structure of the RBM model is presented in Figure 3. The conditions are devoid of connections within themselves. This model is also known as a model based on energy. There exists a multiplicative type of energy between the layers. The model learns by altering this energy variant according to the requirement. The distribution of the probability-based function for the RBM model is denoted as in (1): Here,

( )
, F vw hw represents the multiplicative energy which is as shown in (2): Here, xy w denotes the weight variable, y m and x n denotes the bias variables. The bias variable may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.

Figure 3
Structure of Restricted Boltzmann Machine The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: (1) Here, ( ) , F vw hw represents the multiplicative energy which is as shown in (2): Output: Nodes , , ., in the output layer Step 1: Initialize b as bias and w as weights Step 2: For each weight w xy Step 3: Step 4: For each bias y σ Step 5: Step 6: For each training sample s in N Step 7: For each unit x in the input layer Step 8: Out In Step 9: For each unit y in the hidden layer Step 10: Hid W In b y xy x y = Σ + Step 11: Step 12: For each unit z in the output layer Step 13: Step 14: For each unit y in the hidden layer, moving from the final to initial layer in the hidden unit Step 15:

Figure 2
Structure of Multilayer Perceptron connections within themselves. This model is also known as a model based on energy. There exists a multiplicative type of energy between the layers. The model learns by altering this energy variant according to the requirement. The distribution of the probability-based function for the RBM model is denoted as in (1): Here,

( )
, F vw hw represents the multiplicative energy which is as shown in (2): Here, xy w denotes the weight variable, y m and x n denotes the bias variables. The bias variable may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.

Figure 3
Structure of Restricted Boltzmann Machine The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: (3) (2) Here, xy w denotes the weight variable, y m and x n denotes th e bias variables. The bias variable may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.
The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: or each unit y in the hidden layer, moving inal to initial layer in the hidden unit of Multilayer Perceptron may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.

Structure of Restricted Boltzmann Machine
The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: .
Here, P is considered as a standardization component which is represented as in (4): Here, P is considered as a standardization component which is represented as in (4): In case of two states existence, there are two probabilities calculated assuming that one value is given using the transfer function denoted in (5): The first probability is computed as shown in (6) by assuming that the value of hw is given: Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: . (4) In case of two states existence, there are two probabilities calculated assuming that one value is given using the transfer function denoted in (5): Here, P is considered as a standardization component which is represented as in (4): In case of two states existence, there are two probabilities calculated assuming that one value is given using the transfer function denoted in (5): The first probability is computed as shown in (6) by assuming that the value of hw is given: Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: The weight matrix xy w is modified as shown in (8) to improve the learning of the model: .
The first probability is computed as shown in (6) by assuming that the value of hw is given: Here, P is considered as a standardization component which is represented as in (4): In case of two states existence, there are two probabilities calculated assuming that one value is given using the transfer function denoted in (5): The first probability is computed as shown in (6) by assuming that the value of hw is given: Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: The weight matrix xy w is modified as shown in (8) to improve the learning of the model: . (6) Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: Here, P is considered as a standardization component which is represented as in (4): In case of two states existence, there are two probabilities calculated assuming that one value is given using the transfer function denoted in (5): The first probability is computed as shown in (6) by assuming that the value of hw is given: Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: ( ) The weight matrix xy w is modified as shown in (8) to improve the learning of the model:  1) Here,

 
, F vw hw represents the multiplicative energy which is as shown in (2): Here, xy w denotes the weight variable, y m and x n denotes the bias variables. The bias variable may be modified in order to train the model in a better way to improve the performance. The computational complexity of the system depends on number of layers and techniques used to implement features that are selected.

Structure of Restricted Boltzmann Machine
The probability for the minimal function for the viewable layers is computed as a sum function using Equation (3) as: 439 Information Technology and Control 2023/2/52 The weight matrix xy w is modified as shown in (8) to improve the learning of the model: Similar to this, the second probability is computed as shown in (7) by assuming that the value of vw is given: The weight matrix xy w is modified as shown in (8) to improve the learning of the model:

Deep Belief Networks (DBNs)
Deep belief networks (DBN) are probability distribution models of numerous hidden layers. Adjusting the weights between nodes increases the likelihood that the whole model will be generated. The DBN's fundamental structure is depicted in Figure 4. It consists of a backpropagation (BP) neural network and several restricted Boltzmann machines (RBMs). RBM is a two different neural network with bidirectional connections, with its output being passed as input to the subsequent RBM. Thus, a continuous overlay of an inner layer structure is possible

Figure 4
Structure of Deep Belief Network further updated as shown in (10)  . (12) The above equations, In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.

Deep Belief Networks (DBNs)
Deep belief networks (DBN) are probability distribution models of numerous hidden layers. Adjusting the weights between nodes increases the likelihood that the whole model will be generated. The DBN's fundamental structure is depicted in Figure 4. It consists of a backpropagation (BP) neural network and several restricted Boltzmann machines (RBMs). RBM is a two different neural network with bidirectional connections, with its output being passed as input to the subsequent RBM. Thus, a continuous overlay of an inner layer structure is possible Here, k denotes the length of the data in the training examples. The weight and bias are further updated as shown in (10) The above equations, Here, k denotes the length of the data in the training examples. The weight and bias are further updated as shown in (10)  . (12) . (9) Here, k denotes the length of the data in the training examples. The weight and bias are further updated as shown in (10)  The loss function is optimized on the training data to achieve a nominal solution as shown in (9): Here, k denotes the length of the data in the training examples. The weight and bias are further updated as shown in (10) to (12): The above equations, The above equations, eight matrix xy w is modified as shown in (8) to ve the learning of the model:

eep Belief Networks (DBNs)
belief networks (DBN) are probability bution models of numerous hidden layers. sting the weights between nodes increases the hood that the whole model will be generated. The 's fundamental structure is depicted in Figure 4  . (12) The above equations, . (15) In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.
. (12) The above equations, prob N and prob G denote the numerical and generative probability of the model. The DBN's training strategy is broken up into two phases. Each RBM is first trained from the ground up. The second stage involves top-down fine-tuning of the hyperparameters.
The estimation of the distribution of RBM model is computed mathematically as shown in Equations (13)-(15): ht matrix xy w is modified as shown in (8) to the learning of the model: (8)

Belief Networks (DBNs)
elief networks (DBN) are probability ion models of numerous hidden layers. g the weights between nodes increases the d that the whole model will be generated. The ndamental structure is depicted in Figure 4. It of a backpropagation (BP) neural network and estricted Boltzmann machines (RBMs). RBM different neural network with bidirectional ons, with its output being passed as input to quent RBM. Thus, a continuous overlay of an er structure is possible The above equations, In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.
; (13) ght matrix xy w is modified as shown in (8) to the learning of the model:

Belief Networks (DBNs)
belief networks (DBN) are probability ion models of numerous hidden layers. g the weights between nodes increases the d that the whole model will be generated. The undamental structure is depicted in Figure 4. It of a backpropagation (BP) neural network and estricted Boltzmann machines (RBMs). RBM different neural network with bidirectional ons, with its output being passed as input to equent RBM. Thus, a continuous overlay of an er structure is possible The above equations, In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.
; (14) the learning of the model:

Belief Networks (DBNs)
elief networks (DBN) are probability ion models of numerous hidden layers. g the weights between nodes increases the d that the whole model will be generated. The ndamental structure is depicted in Figure 4. It of a backpropagation (BP) neural network and estricted Boltzmann machines (RBMs). RBM different neural network with bidirectional ons, with its output being passed as input to quent RBM. Thus, a continuous overlay of an er structure is possible The above equations, In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.
. (15) In the above equations, θ denotes the rate of learning by the model and ϕ is the coefficient representing the factor of momentum.

Results and Discussion
For the performance evaluation of the proposed model, experiments have been conducted using two publicly available datasets that comprise ECG signals. The data is preprocessed and fed into the deep hybrid models. A detailed description of the dataset, as well as the experimental findings, are discussed in this section.

Dataset Specification
The two datasets selected for the experimentation of the present research are MIT-BIH [11] and PTB-ECG [20] datasets. The details about the number of samples in the dataset and how it is utilized in this work are presented in Table 3. The total number of samples in the MIT-BIH dataset is 268, whereas in PTB-ECG is 200. Out of 268, 104 samples are for patients affected by cardiovascular disease, and 164 samples represent patients without the condition. Similarly, for the PTB-ECG dataset, the affected patient's count is close to 54, and the unaffected is 146. The dataset is split in the ratio of 75:25 for training and validation.
In that case, for the MIT-BIH dataset, 201 samples are taken for the training phase, and 67 are used for the validation phase. Considering the PTB-ECG dataset, 150 samples are taken for the training phase, and 50 are employed for validation.

Performance Metrics
The performance of the models is evaluated using three different metrics such as  . (17) c) auc model : The relationship between accuracy and precision is plotted on the receiver operating characteristics (ROC) curve. It alludes to a binary classification assessment matrix. The ROC's AUC measures its capacity to differentiate between both positive (16) b . (17) c) auc model : The relationship between accuracy and precision is plotted on the receiver operating characteristics (ROC) curve. It alludes to a binary classification assessment matrix. The ROC's AUC measures its capacity to differentiate between both positive . The relationship between accuracy and precision is plotted on the receiver operating characteristics (ROC) curve. It alludes to a binary classification assessment matrix. The ROC's AUC measures its capacity to differentiate between both positive and negative classifications. The performance of a classifier improves with increasing AUC values.

Performance Evaluation
The performance of the proposed model for both the datasets, such as MIT-BIH and PTB-ECG, is presented in  The performance of the proposed model is compared against various deep learning models, and the accuracy exhibited by each model is presented in Table 5 and graphically depicted in Figure 5.  The proposed hybrid deep network model is also compared against existing hybrid models in [20], [32], [26], [7], and [5] as shown in Table 6. In [20], a hybrid model combining the CNN architectures such as ResNet50, AlexNet, and SqueezeNet has been proposed. In [32], an ensemble model with Deep Neural Networks, CNN, and LSTM is suggested. In [26], another ensemble model based on Self Organizing Maps and Autoencoders is developed for CVD diagnosis. Similarly, in [7], Deep Belief Networks along with XGBOOST are employed for heart disease prediction. In [5], authors have devised a deep learning model MobileNet v2-based DNN and utilized it for CVD research [2,15,19].

Figure 6
Existing Works vs Proposed Work The hybrid model in [20] produces an accuracy of 92.4% for MIT-BIH and 95.8% for PTB-ECG. On the other hand, the ensemble model in [32] exhibits an accuracy of 95.2% and 96.4% for MIT-BIH and PTB-ECG datasets, respectively. The models proposed in [26] produce accurate predictions of 94.3% for MIT-BIH and 96.8% for PTB-ECG datasets. DBN and XGBOOST combination in [7] works well to exhibit forecasts with an accuracy of 93.4% and 95.2% for MIT-BIH and PTB-ECG datasets, respectively. The model in [5] shows a lower accuracy than other models in the existing works taken for performance analysis. It exhibits 91.2% for the MIT-BIH dataset and 92.8% for the PTB-ECG dataset. However, the accuracy of all these models is lower than that of the proposed HybDeepNet model, as shown in Figure 6.  The proposed hybrid deep network model is also compared against existing hybrid models in [20], [32], [26], [7], and [5] as shown in Table 6. In [20], a hybrid model combining the CNN architectures such as ResNet50, AlexNet, and SqueezeNet has been proposed. In [32], an ensemble model with Deep Neural Networks, CNN, and LSTM is suggested. In [26], another ensemble model based on Self Organizing Maps and Autoencoders is developed for CVD diagnosis. Similarly, in [7], Deep Belief Networks along with XGBOOST are employed for heart disease prediction. In [5], authors have devised a deep learning model MobileNet v2-based DNN and utilized it for CVD research [2,15,19].
The hybrid model in [20] produces an accuracy of 92.4% for MIT-BIH and 95.8% for PTB-ECG. On the other hand, the ensemble model in [32] exhibits an accuracy of 95.2% and 96.4% for MIT-BIH and PTB-ECG datasets, respectively. The models proposed in [26] produce accurate predictions of 94.3% for MIT-BIH and 96.8% for PTB-ECG datasets. DBN   another ensemble model based on Self Organizing Maps and Autoencoders is developed for CVD diagnosis. Similarly, in [7], Deep Belief Networks along with XGBOOST are employed for heart disease prediction. In [5], authors have devised a deep learning model MobileNet v2-based DNN and utilized it for CVD research [2,15,19].
The hybrid model in [20] produces an accuracy of 92.4% for MIT-BIH and 95.8% for PTB-ECG. On the other hand, the ensemble model in [32] exhibits an accuracy of 95.2% and 96.4% for MIT-BIH and PTB-ECG datasets, respectively. The models proposed in [26] produce accurate predictions of 94.3% for MIT-BIH and 96.8% for PTB-ECG datasets. DBN and XGBOOST combination in [7] and XGBOOST combination in [7] works well to exhibit forecasts with an accuracy of 93.4% and 95.2% for MIT-BIH and PTB-ECG datasets, respectively. The model in [5] shows a lower accuracy than other models in the existing works taken for performance analysis. It exhibits 91.2% for the MIT-BIH dataset and 92.8% for the PTB-ECG dataset. However, the accuracy of all these models is lower than that of the proposed HybDeepNet model, as shown in Figure 6.

Conclusion
The proposed model, HybDeepNet, utilizes a unique hybrid approach for diagnosing Cardiac Arrhythmia using ECG signals. This includes the use of three different models: Multilayer Perceptrons for optimizing the model's hyperparameters, a Restricted Boltzmann Machine for feature extraction, and Deep Belief Networks for classification. The model is evaluated using two publicly available datasets, MIT-BIH and PTB-ECG. The proposed HybDeepNet model was compared to several deep learning models and two previous works in terms of performance. The results show that the proposed model outperforms the others, achieving an accuracy of 98.6% on the MIT-BIH dataset and 97.1% on the PTB-ECG dataset. However, one limitation of this research is that the analysis was only conducted on consistent datasets. Further validation is needed using larger, more diverse datasets. Additionally, the computational requirements, specifically CPU time, for the hybrid model are relatively high. The proposed HybDeepNet model exhibited superior performance for diagnosing Cardiac Arrhythmia using ECG signals, but there is potential for further improvement. One potential approach is to integrate different types of classification techniques, such as nature-based optimization methods, to enhance the model's performance. Additionally, the same method could be adapted for detecting other diseases by using deep learning detection models.
To aid remote patients, the technique could be implemented on an Internet of Things platform as a cloud service. Another area of focus for future work could be on increasing the operating speed of the model through fast convergence, which would enable real-time diagnosis. Finally, reducing implementation overhead for the developed strategy is another area where work can be done further.