Efficient IoT-based heart disease prediction framework with Weight Updated Trans-Bidirectional Long Short Term Memory-Gated Recurrent Unit

Abstract

The integrated system has generated numerous features for the users, like as identifying heart disease by its symptoms, forwarding the information to the doctors regarding the phase of the probability of disease as well as aiding to fix it. When an emergency situation exists, the system forwards the emergency alert to the respective doctor. Moreover, the automatic system is needed to diagnose heart disease but, the larger data is not sufficient to train the model. Thus, the Internet of Things (IoT) is employed to manage the huge amount of data. Therefore, a novel prediction of heart diseases is implemented with the aid of IoT-based deep learning approaches. Here, the collected data is collected from the three standard databases and then perform preprocessed over the gathered data. Here, the IoT assisted deep learning model is performed to predict heart related diseases accurately. Further, the acquired features of heart diseases are selected using the developed Hybrid Chameleon Electric Fish Swarm Optimization (HCEFSO) via Chameleon Swarm Algorithm (CSA) and Electric Fish Optimization (EFO). Then, the optimally selected features are fed to the training process, where the Trans-Bi-directional Long Short-Term Memory with Gated Recurrent Unit (Trans-Bi-LSTM-GRU) is adopted for predicting heart diseases. Here, the weights are updated with the developed HCEFSO while validating the training phase. The trained Trans-Bi-LSTM-GRU network is used in the testing phase for predicting heart diseases.

Keywords

Hybrid Chameleon Electric Fish Swarm Optimization IoT-based heart disease prediction Gated Recurrent Unit Weight Updated Trans-Bidirectional Long Short Term Memory

1. Introduction

In considering the future generation technologies, the emergence of IoT systems has been required for enhancement [28]. In general, the IoT system has been bounded through various objects that have been invisibly attained from all around the environment [44]. Here, smart Health Monitoring (HM) is demonstrated as the effective computing process to remote the HM along with the IoT system. In addition, it is regarded as the utmost common research application over wearable electronics [2]. Moreover, the Body Sensors Networks (BSN) have functioned as the basic components in the IoT heart disease is a crucial disease [37], which may affect the overall functionality of the heart as well as increase the level of complication [39]. Some of the major symptoms of heart disease are fatigue, weight loss, leg swelling, skin color changes, syncope, dizziness, bradycardia, tachycardia, abdominal pain, neck pain, leg chills, breath shortness, pressure, pain, chest tightness [13]. But, these symptoms may vary under the nature of the heart disease, like dilated cardiomyopathy, mitral regurgitation, congenital heart disease, heart failure, myocardial, and arrhythmia [43].

Moreover, the automatic detection techniques may sometimes acquire a lot of data or may fail to validate the big data, which has created complexity in the system model [3]. Further, this complexity of the system may lead to a decrease in the overall effective rate of the recognition model [33]. So, the automatic system has been combined with the smart devices depending upon the IoT for gathering information about the patients. Here, the system has included the aggregation of sensors as well as devices, which are utilized for attaining the data from the specified environment [40]. Some of the benefits of IoT devices have been harnessed by combining them with the automatic disease recognition model for identifying heart disease. Then, the seriousness of the heart disease requires a screening process for detecting it [12]. During this process, the techniques may detect heart disease features along with the low accuracy rate because of its ineffective validation process, learning, and training [27]. Moreover, the screening process has acquired numerous human interventions and time-consuming manual activities. At the time of screening, the doctors may conduct a stress test, cardiac Computer Tomography (CT), Electrocardiography (ECG), blood glucose level test, and so on. Consequently, the IoT system has generated numerous data related to that health condition and thus attained more attention from the healthcare industries [29].

In addition to that, deep learning-dependent techniques regarding healthcare applications have been highly sophisticated as well as acquired numerous computational aspects for both prediction and training processes [17]. This device has been used in order to gather vital information, such as it refreshes the weightiness over the period and relevant alteration in the well-being restraints [30]. It has also attained more time for training complex neural networks as well as validating the data by utilizing it [35]. Therefore, an intelligent deep learning-based heart disease monitoring system is developed to provide higher detection results [8]. Thus, the heart disease prediction framework with an optimized model has been developed.

Some of the major attribution made in the recommended efficient IoT-based heart disease prediction model is listed below.

To implement novel techniques for IoT-based heart disease prediction models by the Trans-Bi-LSTM-GRU networking model as well as the hybrid HCEFSO model is highly used in the hospital and research field.

To effectively predict heart disease, the newly developed Trans-Bi-LSTM-GRU network model has been developed, and it has been used for the training and testing phase, where the weights of hidden layer in the Trans-Bi-LSTM-GRU network are optimized with the HCEFSO model.

To develop the HCEFSO algorithm via CSA and EFO algorithms that have helped for optimizing the weight in the trained model in Trans-Bi-LSTM-GRU techniques and to update the weight while examining the training phase.

To evaluate the model performance for enhancing the overall performance of the efficient IoT-based heart disease prediction model regarding various metrics with conventional models.

The upcoming section is the proposed model. Phase II gives the literature survey. Deep learning-assisted IoT-based heart disease prediction framework is in phase III, hybrid chameleon electric fish swarm optimization-based optimal feature selection in phase IV, IoT-enabled heart disease prediction by Trans-Bi-LASM-GRU is in phase V, the result is in phase VI, and the conclusion is in phase VII.

2. Literature survey

2.1. Related work

In 2019, Sarkar et al. [1] developed a new methodology, which has been termed as an IOT-dependent Deep Belief Neural Network (DBNN)” to recognize heart-related issues. Further, deep learning techniques can learn heart disease features through attaining the effective manipulation of complex data as well as from past validations. Thus, the IoT-dependent [42] and recommended method validations have effectively reduced heart disease along with reducing heart disease morality by minimizing the complexity of detecting the heart disease. Finally, the validation of the performance has been carried out depending upon the numerous performance metrics.

In 2020, Sarmah et al. [37] designed the IoT that depended on a model for effectively identifying heart disease called “Deep Learning Modified Neural Network (DLMNN)”. In the initial phase, the data related to heart patients have been aggregated. The encrypted data was then decrypted by adopting the proposed classifier model. Then, the categorization was made and thus helps the physician to know the condition of the patients and to treat them. In addition to that, the conventional algorithm in this model has been utilized to secure the transmission data outcomes over the highest level of security as well as to attain the lowest time for encryption.

In 2020, Khan et al. [18] explored the IoT framework for recognizing heart disease by utilizing the model known as “Modified Deep Convolutional Neural Network (MDCNN)”. Thus, the explored model has been used for effectively categorizing the attained sensor data into an abnormal or normal state. The heart monitor and the smartwatch devices were attached to the patients to monitor their condition of the patients. Thus, the outcomes have demonstrated that the explored model, depending upon the heart disease prediction, has outperformed well over other models.

In 2020, Pan et al. [28] implemented a new model for validating heart disease more effectively by utilizing the newly proposed model termed as “Enhanced Deep learning-assisted Convolutional Neural Network” for enhancing the prognostics of patients with heart disease. However, the model was used over the deeper architectural model, which it has covered the multilayer perception techniques along with the regularization of learning techniques. The classifier model has been implemented to aid doctors in effectively detecting heart patients. The validation results have shown that the implemented model was flexible and gained a higher value of precision.

In 2018, Kumar et al. [20] suggested the IoT depended on a scalable “three-tier” framework for detecting heart diseases. In the first tier, the data collection has been carried out through the sensor devices [9,47]. In the second tier, a huge amount of sensor data in cloud computing has been restored. In the third tier, the logistic regression depended on prediction techniques for heart diseases have been explored. In the end, detecting the most relevant clinical parameters for attaining heart disease has been carried out by performing numerous validation metrics.

In 2020, Basumatary et al. [41] demonstrated the new techniques, which were defined as “HealthFog” in order to combine the ensemble techniques over Edge computing devices as well as employed them over real-life applications for automatically detecting heart disease. Moreover, it has generated healthcare for managing the data of heart patients [4,15]. It was then configurable to numerous operation modes that attained better accuracy over diverse fog computation validation as well as for various user requirements.

In 2022, Munagala et al. [26] exploited the novel “Fuzzy-LSTM” techniques for a unique IoT depending on the detection of heart disease framework. Initially, the data was aggregated via wearable IoT devices. Further, the modified algorithm has been used for choosing the most relevant features along with the aid of the objective function. The monitoring system was regarded as the most significant phase for the persistent healthcare mode. Thus, the developed model can minimize the mortality rate, healthcare monitoring techniques have been developed. Hence, the validation has proved that the exploited method has successfully predicted heart illness.

In 2022, XinShi et al. [38] proposed a new methodology named “Attention Mechanism-Enabled Bi-Directional-LSTM (ABi-LSTM) techniques depending upon the ECG signals for automatically detecting heart disease. Here, the ECG signal data has been attained through two publically available resources for both testing and testing the newly proposed model. These techniques have made the model features have the ability to record the information from the signals along with the adaptive learning model. In addition to that, the complex features were also retrieved from the ECG signals and performed the detection process. Thus, the experiment has shown the highest rate of accuracy level for the proposed model. Hence, the recommended technique has provided better outcomes.

Table 1
Benefits and drawbacks of existing heart disease prediction models using deep learning techniques.

Author [citation] Techniques Benefits Drawbacks

Makhadmeh et al. [1] HOBDBNN
It provides early diagnosis and treatment.

It increases the efficiency in the effective manipulation of complex data.

It does not use any optimization algorithm.

The security of the IoT data was not considered.

Sarmah et al. [37] DLMNN
It takes the lowest time for encryption.

This model contains the highest level of the security system.

It requires a high cost for implementation.

Khan et al. [18] MDCNN
The false classification rates are low in this model.

It reduces the time complexity of data.

The prediction of precision measure value is very low.

Pan et al. [28] EDCNN
It reduces heart disease mortality.

It contains higher modeling capacity parameters for efficient training.

This model does not select the algorithm for identifying the best mobile ambulance for analyzing heart disease data.

Limited training data is available so the EDCNN performance is very low.

Kumar et al. [20] Scalable three-tier(ML)
It decreases the number of misdiagnoses.

Using an IoT-based framework to increase scalability and availability.

More resources are required.

Basumatary et al. [41] Deep learning
It achieves high accuracy with very low latencies.

It supports both global and public health.

The cost of effectiveness is very high.

A lack of technical knowledge is needed.

Munagala et al. [26] Fuzzy-LSTM
It is flexible to train complex problems.

It provides low power consumption in real-time.

It is unable to manage high-dimensional information and also causes low scalability and efficiency.

XinShi et al. [38] ABLSTM
It validates the heart patient data in real-time and reduces the risk to the patients.

It increases the accuracy prediction value and precision value.

A lot of memory is required.

It requires more time for training.

Author [citation]	Techniques	Benefits	Drawbacks
Makhadmeh et al. [1]	HOBDBNN	It provides early diagnosis and treatment. It increases the efficiency in the effective manipulation of complex data.	It does not use any optimization algorithm. The security of the IoT data was not considered.
Sarmah et al. [37]	DLMNN	It takes the lowest time for encryption. This model contains the highest level of the security system.	It requires a high cost for implementation.
Khan et al. [18]	MDCNN	The false classification rates are low in this model. It reduces the time complexity of data.	The prediction of precision measure value is very low.
Pan et al. [28]	EDCNN	It reduces heart disease mortality. It contains higher modeling capacity parameters for efficient training.	This model does not select the algorithm for identifying the best mobile ambulance for analyzing heart disease data. Limited training data is available so the EDCNN performance is very low.
Kumar et al. [20]	Scalable three-tier(ML)	It decreases the number of misdiagnoses. Using an IoT-based framework to increase scalability and availability.	More resources are required.
Basumatary et al. [41]	Deep learning	It achieves high accuracy with very low latencies. It supports both global and public health.	The cost of effectiveness is very high. A lack of technical knowledge is needed.
Munagala et al. [26]	Fuzzy-LSTM	It is flexible to train complex problems. It provides low power consumption in real-time.	It is unable to manage high-dimensional information and also causes low scalability and efficiency.
XinShi et al. [38]	ABLSTM	It validates the heart patient data in real-time and reduces the risk to the patients. It increases the accuracy prediction value and precision value.	A lot of memory is required. It requires more time for training.

2.2. Problem specifications

Most of heart disease detection model demands, high-performance hardware consisting of multi-core graphics processing units that require a lot of electricity and a lot of computational time in the early stage. Therefore, many deep learning techniques have been developed for heart disease prediction, and some of them are listed in Table 1. HOBDBNN [1] provides early diagnosis and treatment. Also, it increases the efficiency in the effective manipulation of complex data. But, it does not use any optimization algorithm. DLMNN [37] takes the lowest time for encryption against the existent AES. Also, this model contains the highest level of the security system. Yet, requires a high cost for the implementation. MDCNN [18] it ensures a lower value of prediction of precision measure value. EDCNN [28] reduces heart disease mortality. Also, it contains higher modeling capacity parameters for efficient training. But, the model does not select the algorithm for identifying the best mobile ambulance for analyzing the heart disease data, and also limited training data is available, so the EDCNN performance is very low. Scalable three-tier (ML) [20] decreases the number of misdiagnoses. Using an IoT-based framework is utilized to increase scalability and availability. Yet, more resources are required, and also the errors are not corrected by the initial stage. Deep learning [41] achieves high accuracy with very low latencies. Also, supports both global and public health. But, the cost of the effectiveness is very high, and also a lack of technical knowledge is needed. Fuzzy-LSTM [26] is flexible to training complex problems. Also, provides low power consumption to monitor patients in real-time. Yet, it is unable to manage high-dimensional information, and also it causes low scalability and efficiency. ABLSTM [38] validates the heart patient data and reduces the risk to the patients. Also, it increases the accuracy prediction value and precision value. But, a lot of memory is required.

Motivation

By resolving the above mentioned critical issues a new IoT-based heart disease prediction model is implemented using intelligent and machine learning heuristic algorithms. Somehow, heart disease prediction is popularly utilized in real time applications like healthcare and mobile based applications. Additionally, the research work is focused to implement a novel method called Trans-Bi-LSTM-GRU. The overfitting issues are resolved by the developed method, and also it has the tendency to solve the gradient vanishing problems. Often, the designed HCEFSO algorithm is suggested for parameter optimization among the classifiers. Thus, it helps to fasten the performance of the developed model. However, the error rate gets minimized to provide effective outcomes. Here, the validation shows that the offered method provides better performance than the existing methods. Thus, it has the ability to identify heart disease in an accurate manner and also doctors can easily diagnose the disease.

3. Deep learning-assisted IoT-based heart disease prediction framework

3.1. IoT-based heart disease prediction

In this modern era, most hospitals have maintained health care data through health information technology, in which the technology has included a huge amount of data that are utilized for retrieving the hidden information for making an intelligent medical diagnosis. On considering machine learning techniques, it has the potential to process huge amounts of data as well as reliably transformed clinical data, which has aided physicians in providing better detection accuracy. In addition to that, an automatic alert generation feature has also been developed in the case of an emergency. Here, IoT-based techniques have been considered a vast field [16,23]. Some of the models like Random Forest, Decisions Tree, and Neural Networks are utilized in the detection process. Consequently, deep learning techniques have been used in healthcare applications [19,32] that take more time for training the complicated neural networking model as well as validate the data. Some of the conventional techniques have also suffered through complexity in IoT applications and healthcare over real-time applications [6,10]. Further, it requires high time for prediction and huge computational resources for both the detection and training process [32]. Hence, there is an emergency requirement for an automatic heart disease detection model using an IoT framework [21,31,45]. The architectural representation of the designed IoT-based heart disease system is illustrated in Fig. 1.

Fig. 1.

Efficient IoT-based heart disease prediction framework architectural model.

The proposed model is carried out in three different phases. The required data has been aggregated initially, and then the preprocessing of raw data has been performed over the aggregated data. Then, the relevant features of heart diseases have been chosen by utilizing the HCEFSO algorithm model that has been designed via CSA and EFO. Further, the optimally selected features have been forwarded to the training process, in which the technique termed Trans-Bi-LSTM-GRU is adopted to predict heart diseases. Here, the weights of the hidden layer in the Trans-Bi-LSTM-GRU have been updated with the aid of the HCEFSO algorithm while validating the trained model. The trained Trans-Bi-LSTM-GRU network has been utilized in the testing phase for the prediction of heart diseases. Finally, the experimental results have been assimilated by different conventional IoT-based heart disease prediction systems for testing the effectiveness of the given model.

3.2. Datasets of collected heart disease prediction

The data that are significantly required to perform the efficient IoT-based heart disease prediction framework has been initially gathered from three datasets. The dataset links and its description are detailed in Table 2.

Table 2
Datasets description in the heart disease prediction models and their description.

Dataset Link Description

Dataset 1 “https://archive.ics.uci.edu/ml/datasets/Heart+failure+clinical+records”: “Access date: 2023-01-30” It is named “Heart failure clinical records Data Set”. Moreover, the information is are taken from the “Government College University, Faisalabad, Pakistan”. This particular dataset contains 12 attributes, which involve serum sodium, time, age, sex, smoking, serum creatinine, high blood pressure, creatinine phosphokinase, platelets, anaemia Feature Binary, diabetes, and ejection_fraction

Dataset 2 “https://www.kaggle.com/ronitf/heart-disease-uci”: “Access date: 2023-01-30” It is demonstrated as the Heart Disease Data Set from the UCI data repository. It has involved a total of 14 subsets. The major task included in this dataset is to effectively predict the disease depending upon the provided attributes of patients, whether the particular person has heart disease or not. It has also attained experimental tasks for detection issues. Further, it has acquired 16 columns. Here, several variables like age, sex (1 belongs to male and 0 belongs to female), fast blood sugar, resting electrocardiographic, chest pain type, which includes (typical angina = 0, atypical angina = 1, non-anginal pain = 2, and asymptomatic = 3), cholesterol, maximum heart rate, exercise induced angina, and also the slope of the peak exercise ST segment (upsloping = 0, flat = 1, downsloping = 2).

Dataset 3 “https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset”: “Access date: 2023-01-30” It is termed a Cardiovascular Disease dataset. It has included 12 columns. There is a total of 70 000 data records from the patients are included. Further, 11 or more targets as features are used in this dataset. It has three types of input features such as information given by the patient, results of medical examination, and factual information. In dataset 3, the variables like age, cholesterol, height, gender, Smoking, weight, Alcohol intake, Glucose, and Physical activity.

Dataset	Link	Description
Dataset 1	“https://archive.ics.uci.edu/ml/datasets/Heart+failure+clinical+records”: “Access date: 2023-01-30”	It is named “Heart failure clinical records Data Set”. Moreover, the information is are taken from the “Government College University, Faisalabad, Pakistan”. This particular dataset contains 12 attributes, which involve serum sodium, time, age, sex, smoking, serum creatinine, high blood pressure, creatinine phosphokinase, platelets, anaemia Feature Binary, diabetes, and ejection_fraction
Dataset 2	“https://www.kaggle.com/ronitf/heart-disease-uci”: “Access date: 2023-01-30”	It is demonstrated as the Heart Disease Data Set from the UCI data repository. It has involved a total of 14 subsets. The major task included in this dataset is to effectively predict the disease depending upon the provided attributes of patients, whether the particular person has heart disease or not. It has also attained experimental tasks for detection issues. Further, it has acquired 16 columns. Here, several variables like age, sex (1 belongs to male and 0 belongs to female), fast blood sugar, resting electrocardiographic, chest pain type, which includes (typical angina = 0, atypical angina = 1, non-anginal pain = 2, and asymptomatic = 3), cholesterol, maximum heart rate, exercise induced angina, and also the slope of the peak exercise ST segment (upsloping = 0, flat = 1, downsloping = 2).
Dataset 3	“https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset”: “Access date: 2023-01-30”	It is termed a Cardiovascular Disease dataset. It has included 12 columns. There is a total of 70 000 data records from the patients are included. Further, 11 or more targets as features are used in this dataset. It has three types of input features such as information given by the patient, results of medical examination, and factual information. In dataset 3, the variables like age, cholesterol, height, gender, Smoking, weight, Alcohol intake, Glucose, and Physical activity.

The data that is aggregated for further performing the efficient IoT-based heart disease prediction framework is indicated as ${H D}_{k}$ , $k = 1, 2, \dots, K$ where the total number of aggregated data is termed as K.

3.3. Raw data pre-processing

In this pre-processing phase, ${H D}_{k}$ , which are aggregated from the initial phase are forwarded. The pre-processing step is regarded as the most significant phase in the data mining process. Here, the process such as transforming, cleaning, and combining data to pre-preparing it for validation. This model aims to enhance the data quality as well as it has been regarded as the most relevant technique. Thus, the pre-processing of data has been utilized it neglects the inconsistent or missing data values. Thus, it has enhanced the quality as well as the accuracy values of the datasets. It has made the data consistent and more reliable. In this efficient IoT-based heart disease prediction framework, the pre-processing of data is made by replacing the Not a Number (NAN) values with zero. It is defined as the general technique for depicting the missing values in the data. Thus, the NAN is regarded as the major issue while performing the validation process. Because it is demonstrated as a special floating-point value as well as it does not transfer to any other type than float. Hence, there is a requirement of dealing with the NAN values for attaining better outcomes. So, the NAN value is replaced in this model by using the zero value for attaining better outcomes. Thus, the pre-processed data attained through this phase has been depicted as $p_{NZ}$ .

4. Hybrid Chameleon Electric Fish Swarm Optimization-based optimal feature selection for heart disease prediction

4.1. Proposed HCEFSO

The newly implemented HCEFSO algorithm model has been developed by using two algorithms like, CSA [7] and EFO [46] algorithm model. Here, the EFO algorithm has been used to solve the problems like convergence as well as determined a balance among exploitation and exploration phases. But, it takes more time for execution which degrades the system performance. Additionally, the CSA is used to enhance the flexibility of the model and it also provides better optimal solutions. But it has faced over-fitting issues. In order to tackle the disadvantages of the conventional model and to enhance the performance of the overall model, new techniques called HCEFSO have been developed.

\begin{aligned} e = \frac{o j f_{bst}}{o j f_{wrs}} \end{aligned}

(1)

Here, the term ${o j f}_{bst}$ , ${o j f}_{wrs}$ depicts the best and worst fitness functions. In this case, if $e > r n d$ then, the update has been made through CSA else the update is made through the EFO algorithm.

CSA is implemented through the dynamic behavior of the chameleons while hunting for their food as well as on navigating over the near swamps, deserts, and trees. It changes its location when roaming over the trees and deserts looking for its food. In the initial phase, the CSA has been defined as the population dependent on the algorithm; it applies the initial population for initiating the process of optimization. Here, the population of the chameleons is indicated as y over the dimensional searching space termed as x along with each chameleon has defined as a candidate solution to issues has been depicted over the two-dimensional matrix a of size denoted as $x \times y$ . Moreover, the location of the chameleons is indicated as c at b iterations in the searching domain has been depicted through the vector and it is equated in Eq. (2).

\begin{aligned} a_{b}^{c} = [a_{b, 1}^{c}, a_{b, 2}^{c}, \dots, a_{b, z}^{c}] \end{aligned}

(2)

Here, the position of the $c^{th}$ chameleons over the $b^{th}$ dimension is indicated as $a_{b, z}^{c}$ , the dimension of the problem is termed as x, and the current iteration number is depicted as $c = 1, 2, \dots, x$ .

The initial population has been developed on the dimension of the problem as well as the number of chameleons in the search space as given in Eq. (3).

\begin{aligned} a^{c} = l w_{d} + e \times (u p_{d} - l w_{d}) \end{aligned}

(3)

Here, the uniformly created random number lies in the interval of $[0, 1]$ is indicated as e, the upper and lower bounds of the search area in the $d^{th}$ dimension are termed as ${u p}_{d}$ , and ${l w}_{d}$ accordingly, the initial vector of the $c^{th}$ chameleon is given as $a^{c}$ .

Hunting prey: The chameleon has completed its hunting of their prey while their prey is very close to the chameleon. The chameleon has utilized its tongue for attacking the prey. The chameleons, which are close to their prey have been supposed as the better chameleons. It is defined as an optimal solution as well. Therefore, its location has been upgraded slightly as it has dropped its tongue to double as its length. Further, this approach has aided the chameleons in exploiting the search space by effectively grabbing prey.

The velocity of the chameleon’s tongue, while it has fallen toward prey has been equated in Eq. (4).

\begin{aligned} f_{b + 1}^{c, d} = ξ f_{b}^{c, d} + g_{1} (A_{b}^{d} - a_{b}^{c, d}) e_{1} + g_{2} (B_{b}^{e, d} - a_{b}^{c, d}) e_{1} \end{aligned}

(4)

Here, the $c^{th}$ chameleon’s best-known position is depicted as $B_{b}^{e, d}$ , two random numbers from 0 to 1 are indicated $e_{1}$ , $e_{2}$ , the best global position of chameleons is given as $A_{b}^{d}$ , and the current position $c^{th}$ chameleon is represented as $f_{b}^{c, d}$ , the new velocity of the $c^{th}$ chameleons over the $d^{th}$ dimension at the iteration $b + 1$ is given as $f_{b + 1}^{c, d}$ two positive constants that control $B_{b}^{e, d}$ and $A_{b}^{d}$ is termed as $g_{1}$ , $g_{2}$ ,as well as the inertia weight, which is linearly minimized has been given as ξ in Eq. (5).

\begin{aligned} ξ = {(1 - \frac{b}{C})}^{(ω \sqrt{\frac{B}{c}})} \end{aligned}

(5)

Here, the positive number has been utilized to determine the exploitation capacity is given as ω.

While enhancing the convergence behavior in the CSA has been initialized with the weight parameters. The location of the chameleon’s tongue that has been computed over the motion is equated in Eq. (6).

\begin{aligned} a_{b}^{c, d} = a_{b}^{c, d} + \frac{((f_{b}^{c, d})^{2} - {(f_{b - 1}^{c, d})}^{2})}{(2 h)} \end{aligned}

(6)

Here, the earlier speed of the $c^{th}$ chameleons over the $d^{th}$ dimension is given as $f_{b - 1}^{c, d}$ , the acceleration rate of the chameleon’s tongue until attaining the maximum value is given as h. It is equated in Eq. (7).

\begin{aligned} h = 2590 \times (1 - w^{- \log (b)}) \end{aligned}

(7)

Consequently, the EFO algorithm is demonstrated as the nature-dependent algorithm model that has been implemented by the prey location of the electric fishes as well as by communication mannerisms.

Passive electro-location: Generally, it has the potential for global searching as well as for the exploration mechanism over the EFO algorithms. In considering the passive phase, the individuals have selected other individuals over the active phase, in which it has shared the electrical signal depending upon the probability and its transformation in the locations.

With regard to the active mode, the probability for the individual termed as $i \in C_{D}$ , in which it has been perceived by the individual indicated as $j \in C_{E}$ over the passive model, which is equated in Eq. (8).

\begin{aligned} E_{i} = \frac{D_{i} / c c_{j i}}{\sum_{z \in C_{D}} D_{z} / v_{j z}} \end{aligned}

(8)

Here, the amplitude has become the dominant factor, which it has caused the local search space and choosing of the best individuals. It has been demonstrated, where the probabilistic neighbor selection has been forced while performing the exploration phase before the exploitation stage.

Here, the individual indicated as Z has been selected by $C_{D}$ , and the reference location $Y_{x x}^{z z}$ is depicted in Eq. (9).

\begin{aligned} Y_{x x}^{z z} & = \frac{\sum_{x x = 1}^{Z} D_{i} Y_{x x}^{z z}}{\sum_{x x = 1}^{Z} D_{i}} \end{aligned}

(9)

\begin{aligned} Y_{x x}^{new} & = Y_{j z} + ζ (Y_{x x}^{z z} - x x_{j z}) \end{aligned}

(10)

Here, the individual has lost the information about the location completely, in which does not accept the particular region that is positioned. For tackling the problems, the EFO algorithm has determined the modified parameters, and it is given in Eq. (10). The probability for such the individual altering the entire trait that is significantly minimized within the acceptance condition and it is given in Eq. (11).

\begin{aligned} Y_{x x}^{vh} = {\begin{cases} Y_{x x}^{new} & r n d_{z} (0, 1) > b b_{j} \\ Y_{x x}^{new} & else \end{cases} \end{aligned}

(11)

Here, the term ${r n d}_{z} (0, 1)$ is indicated as a random number.

Finally, the passive electro-location has altered one of its parameters among the individual indicated as $j^{th}$ that has been used in Eq. (12) for improving the probability of the trait has been transformed.

\begin{aligned} Y_{x x}^{vh} & = Y_{min j} + ζ (Y_{max j} - Y_{min j}) \\ r n d (0, 1) ⩽ r n d (0, 1) \end{aligned}

(12)

Here, the term $r n d (0, 1)$ denotes the random number.

While the $z^{th}$ parameter values over the $j^{th}$ individual space are exceeded.

\begin{aligned} Y_{x x}^{vh} {\begin{cases} Y_{min j} & Y_{x x}^{vh} < Y_{min j} \\ Y_{x x}^{vh} & Y_{max j} > Y_{x x}^{vh} > Y_{min j} \\ Y_{max j} & Y_{x x}^{vh} > Y_{max j} \end{cases} \end{aligned}

(13)

Finally, the overall process repeats until attaining the best optimal solution, and its pseudo-code is given in Algorithm 1.

4.2. HCEFSO-based optimal feature selection

The input is given as $p_{NZ}$ . Here, the most relevant features are utilized to select for further processing phase related to patients data have been optimally selected. Moreover, the HCEFSO algorithm has been used in this phase for optimally selecting the feature, which may help to speed up the overall process of the given efficient IoT-based heart disease prediction model. In addition to that, an objective function has also been used for maximizing the variance of the model, and the first objective function is in Eq. (14).

\begin{aligned} o b j c_{1} = \underset{{o f^{l}}}{\arg min} (\frac{1}{v r n}) \end{aligned}

(14)

Here, the term $v r n$ defines the maximized variance, and it is equated as in Eq. (15).

\begin{aligned} v r n = \frac{\sum_{y z = 1}^{t} {(O_{y z} - \bar{O})}^{2}}{s} \end{aligned}

(15)

Here, $O_{y z}$ and $\bar{O}$ depicts the $y z^{th}$ pixel and “mean value of all pixels in the image”.

During this process, the optimally selected feature data by using the HCEFSO algorithm has been given as ${o f}^{l} l = 1, 2, \dots, L$ is the total number of optimally selected features. The model for depicting the HCEFSO-based optimal feature selection has been given in Fig. 2.

Fig. 2.

HCEFSO-based optimal feature selection for IoT-based heart disease prediction framework.

Fig. 3.

Proposed Trans-Bi-LSTM-GRU model with parameter optimization for the given model.

Fig. 4.

Convergence analysis for the efficient IoT-based heart disease prediction framework regarding “a) dataset 1, b) dataset 2, and c) dataset 3”.

Fig. 5.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 1 with various algorithms regarding “a) accuracy, b) F1-score, and c) precision”.

Fig. 6.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 2 with various algorithms regarding “a) accuracy, b) F1-score, and c) precision”.

Fig. 7.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 3 with various algorithms regarding “a) accuracy, b) F1-score, and c) precision”.

Fig. 8.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 1 with various classifiers regarding “a) accuracy, b) F1-score, and c) precision”.

Fig. 9.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 2 with various classifiers regarding “a) accuracy, b) F1-score, and c) precision”.

Fig. 10.

Analyzing the performance for the efficient IoT-based heart disease prediction framework for dataset 3 with various classifiers regarding “a) accuracy, b) F1-score, and c) precision”.

Table 3

Statistical analysis over dataset 1 to 3 for the efficient IoT-based heart disease prediction framework.

Algorithms	WOA-Trans-Bi-LSTM-GRU [25]	HBA-Trans-Bi-LSTM-GRU [14]	CSA-Trans-Bi-LSTM-GRU [7]	EFO-Trans-Bi-LSTM-GRU [46]	HCEFSO-Trans-Bi-LSTM-GRU
Dataset 1
“Best”	1.002546	1.001448	1.003531	1.000291	1.000173
“Worst”	1.003969	1.05422	1.010446	1.069622	1.080222
“Mean”	1.002997	1.006477	1.004554	1.001703	1.003447
“Median”	1.002546	1.001448	1.003531	1.000322	1.000173
“STD”	0.000658	0.010177	0.001404	0.009703	0.011555
Dataset 2
“Best”	1.001931	1.000617	1.000065	1.000776	1.000039
“Worst”	1.00824	1.021114	1.007311	1.012754	1.039603
“Mean”	1.005989	1.003868	1.000945	1.00151	1.002965
“Median”	1.00824	1.002631	1.000065	1.001222	1.000039
“STD”	0.002742	0.005535	0.002351	0.002305	0.007638
Dataset 3
“Best”	1.000735	1.000703	1.001794	1.001135	1.000031
“Worst”	1.0112	1.153845	1.025908	1.069864	1.071321
“Mean”	1.001884	1.00801	1.00683	1.004846	1.000761
“Median”	1.000735	1.001113	1.007401	1.001135	1.000031
“STD”	0.00241	0.024525	0.005574	0.013483	0.007093

5. IoT-enabled heart disease prediction by Trans-Bi-directional Long Short Term Memory with Gated Recurrent Unit

5.1. Trans-Bi-directional Long Short-Term Memory

Several deep networks have been emerged and progressed in the field of healthcare. However, the heart disease prediction is necessary to decrease the mortality rate. In existing, the long term dependency cannot be captured in the CNN model. In this research work, we have considered deep networks like Trans-Bi-directional Long Short-Term Memory and GRU. Both networks have a stronger capability to predict heart related disease. Considering Trans-Bi-LSTM, it has the ability to fuse several hidden layers, and thus, it solves various complex problems to train the network more effectively. Here, the training takes place towards both forward and backward directions. On the other hand, the GRU is the simplest model that supports handling the long term sequence, and also it carries long term memories. Here, the GRU model provides the lesser processing time to train the model. The optimally selected feature ${o f}^{l}$ has been given as the input to the newly developed Trans-Bi-LSTM-GRU model for training the data and to attain the trained model as outcomes.

The transformer-based techniques over the Bi-LSTM [24] model have been effectively defined as the Trans-Bi-LSTM framework. In this case, the self-attention mechanism has been employed by using the transformer model for carrying out their job. Moreover, it has made a specified progression in using the computer vision-dependent framework. It has also been used for language process regarding the general process. For designing the Trans-Bi-LSTM framework effectively both the encoder as well as the decoder has been integrated into a structure. Further, it has been utilized for executing the sequence-to-sequence task. Here, the encoder block has been used alone in order to determine the outcomes. It has then resulted in the encoded output being fed to the Bi-LSTM techniques.

The optimal selected features are given as input to the transformer approach. Here, the primary function of the transformer model is to compute the attention level. It has been then followed through the determined the attention score along with the help of the “scaled dot product attention model.” Through the input, “the query vector termed as $q v$ key vector indicated as $k v$ as well as the value vector indicated as $v v$ ” is determined by the weight matrices multiplication is depicted as $w g t$ . It is then equated in Eq. (16).

\begin{aligned} mh (o f^{l}, o f^{l}, o f^{l}) = ct (L_{1}, \dots, L_{b a}) w t \end{aligned}

(16)

Here, the ${q q}^{th}$ head is depicted as $L_{q q}$ that is expressed as in Eq. (17).

\begin{aligned} as (o f^{l q v}, o f^{l k v}, o f^{l v v}) = smx (\frac{o f^{l q v} ∙ {(o f^{l v q})}^{AGA}}{\sqrt{n e_{bv}}}) w t \end{aligned}

(17)

Thus, the results attained from the transformer are taken as ${L T}^{nec}$ , and it is then fed as input to the Bi-LSTM model.

In an LSTM network is defined as the type of RNN technique in which it has utilized four gates. This gate can decide which portion of the detail is needed to be propagated or sent or forgotten. Here, the LSTM model also has the potential to outperform well in applications, which all need to be carried out in the long sequence of data when assimilated over the RNN model, as well as minimized the over-fitting issues in the RNN model. It technique has trained two LSTMs on the input sequence. Among these two, one of the input sequences through the first step data to the last one, as well as the second over the reversed copy of the input layer. Thus, in this case, it has provided an additional context to the network as well as the outcomes is faster.

Here, the Bi-LSTM techniques have processed the given data in both the backward and forward orientation by two separate LSTM layers. Further, the forward hidden state has been depicted as ${\vec{M}}_{i i}^{g g}$ as well as the backward state indicated as ${\overset{\leftrightarrow}{M}}_{i i}^{g g}$ has been determined and consequently are concatenated, and then it has been subjected or forwarded to the next layer is given in Eq. (18).

\begin{aligned} {\overset{\leftrightarrow}{M}}_{i i}^{g g} = [\begin{matrix} {\vec{M}}_{i i}^{g g} \\ {\overset{\leftarrow}{M}}_{i i}^{g g} \end{matrix}] \end{aligned}

(18)

Here, the input layer is represented as $g g = 0$ .

5.2. Gated Recurrent Unit

The GRU [48] techniques have been included in the RNN model. The RNN model has the potential to memorize the arbitrary length of the input pattern by developing a connection among the units through the directed cycle. The major key component over the RNN has been defined as the transition function in every time step indicated as $t m$ that carried out the current time information denoted as $F_{t m}$ as well as the previously hidden layer has been depicted as $G_{t m - 1}$ and it is used to update the current hidden output layer and equated in Eq. (19).

\begin{aligned} G_{t m} = H (F_{t m}, G_{t m - 1}) \end{aligned}

(19)

Here, the differentiable and the nonlinear transformation function are indicated as H. The RNN technique holds its memory of the previous input over the internal state of the networking model, and thus, the recurrent structure has been considered as the memory of previous inputs. A supervised learning layer has been added on top to map the attained depictions $G_{t m}$ to targets as well as the techniques have been trained through backpropagation via time. Various formulations in the transition function have been used to define the RNN model. The RNN model has acquired the linear transformation function is expressed as in Eq. (20).

\begin{aligned} G_{t m} = ζ (I F_{t m} + J G_{t m - 1} + b b) \end{aligned}

(20)

Here, the bias vector has been depicted as $b b \in ℜ^{ee}$ , and the transformation matrices are indicated as $J \in ℜ^{ee \times ee}$ and $I \in ℜ^{ee \times ff}$ . The non-linear activation functions like tanh and sigmoid functions are represented as ζ. Consequently, the RNN model has failed to capture the long-term dependency due to the vanishing gradient issues that occur at the time of backpropagation for training the model. Thus, the GRU with two gates is developed. The update gate indicated as U are designed. On the other hand, the reset gate denoted as R has been utilized to adjust the incorporation of new input along with the previous memory. The transition function over the hidden unit of the GRU model is expressed in Eq. (21) to Eq. (24).

\begin{aligned} U_{tm} & = ς (I^{U} F_{tm} + J^{U} G_{tm - 1} + b b^{U}) \end{aligned}

(21)

\begin{aligned} R_{tm} & = ς (I^{R} F_{tm} + J^{R} G_{tm - 1} + b b^{R}) \end{aligned}

(22)

\begin{aligned} {\bar{G}}_{tm} & = \tanh (I^{X} F_{tm} + J^{X} (R_{tm} ∙ G_{tm - 1})) \end{aligned}

(23)

\begin{aligned} G_{tm} & = (1 - U_{tm}) ∙ G_{tm - 1} + U_{tm} ∙ {\bar{G}}_{tm} \end{aligned}

(24)

Here, the hyper-parameter that represents the dimensionality of hidden vectors is termed as X, the element-wise product is denoted as •, the terms that are shared by all time steps are indicated as $J \in ℜ^{ee \times ee}$ , $I \in ℜ^{ee \times ff}$ and $b b \in ℜ^{ee}$ .

5.3. Proposed Bi-LSTM-GRU with parameter optimization

Finally, for attained the predicted outcomes, the proposed model has been carried out in three different phases training, validating, and testing. In phase 1, the optimally selected feature ${o f}^{l}$ has been given as the input to the newly developed Trans-Bi-LSTM-GRU model for attaining the trained model. In phase 2, the weights of the hidden layer in the trained model of Trans-Bi-LSTM-GRU are optimized with the aid of HCEFSO while validating the performance of the trained model. In phase 3, the attained trained Trans-Bi-LSTM-GRU network has been utilized for attaining the effective prediction of heart diseases in the testing phase. Final predicted outcomes have been attained in a sequence manner from the Trans-Bi-LSTM-GRU model.

The Bi-LSTM model has been considered as the most beneficial technique that has been used for processes like translation, recognition, and classification. But, this model has been regarded as slower as well as acquires more time for training the data. Consequently, the GRU model has utilized less memory space for processing the data, and it also can provide more accurate accuracy values. It is also used to address the vanishing gradient issues. But, it has a low learning efficiency rate and slow convergence rate. For overcoming the limitation in the conventional model and to further enhance the performance of the prediction model, the newly developed Bi-LSTM-GRU has been implemented to provide a better accuracy rate as well. The objective function is also used to attain the maximized accuracy, and it is given in Eq. (25).

\begin{aligned} o b j c_{1} = \underset{{w g t}}{\arg min} (\frac{1}{a c t r}) \end{aligned}

(25)

Here, the weight among the range $[- 0.06703135, 1]$ has been depicted as $w g t$ in the trained model is optimized with the HCEFSO model for developing the performance of the model. The term $a c t r$ defines accuracy and it is in Eq. (26).

\begin{aligned} a c t r = \frac{(o p + o q)}{(o p + o q + r v + r k)} \end{aligned}

(26)

The terms are representing $o p$ , $o q$ , $r k$ , $r v$ the “true positives, true negatives, false positives, and false negatives”, accordingly. The diagrammatic depiction of the developed Trans-Bi-LSTM-GRU model with parameter optimization for the prediction model is given in Fig.3.

6. Results

6.1. Simulation setup

The implemented IoT-based heart disease prediction model was performed in Python. However, the Python works on various platforms like Windows, Mac, etc. Additionally, Python is simple but it is effectively perform in object oriented programming. The processor Intelcore i3 is taken, and also the RAM size is 8 GB and 64 bit. Additionally, the version is considered as a community version. The “population size was 10, chromosome length was 8, and the maximum number of iterations was 25” were used. Some of the algorithms like “Whale Optimization Algorithm (WOA) [25], Honey Badger Algorithm (HBA) [14], CSA [7] and EFO [46] and classifiers like BILSTM [24], Fuzzy-LSTM [26] CNN-Bi-LSTM [24,34]” have been utilized to evaluate the performance of the developed model.

6.2. Performance metrics

The implemented proposed IoT-based heart disease prediction framework is measured below.

(a)
Sensitivity is in Eq. (27)
$\begin{aligned} s i v = \frac{o p}{o p + r k} \end{aligned}$
(27)

(b)
Specificity is in Eq. (28)
$\begin{aligned} s p c = \frac{o q}{o q + r v} \end{aligned}$
(28)

(c)
FPR is in Eq. (29)
$\begin{aligned} f f r = \frac{r v}{r v + o q} \end{aligned}$
(29)

(d)
FNR is in Eq. (30)
$\begin{aligned} f r = \frac{r k}{o q + o p} \end{aligned}$
(30)

(e)
NPV is in Eq. (31)
$\begin{aligned} n v = \frac{r k}{r k + o q} \end{aligned}$
(31)

(f)
FDR is in Eq. (32)
$\begin{aligned} f d r = \frac{r v}{r v + o p} \end{aligned}$
(32)

(g)
Precision is in Eq. (33)
$\begin{aligned} p s c n = \frac{o p}{o p + r v} \end{aligned}$
(33)

(h)
F1 score is in Eq. (34)
$\begin{aligned} f c s 1 = \frac{s i v \cdot p s c n}{p s c n + s i v} \end{aligned}$
(34)

(i)
MCC is in Eq. (35)
$\begin{aligned} m c c c \frac{o p \times o q - r v \times r k}{\sqrt{(o p + r v) (o p + b a) (o q + r k) (o q + r k)}} \end{aligned}$
(35)

6.3. Convergence validation for the proposed model for dataset 1 to 3

Here, Fig. 4 has depicted the convergence analysis of the proposed model for datasets 1 to 3 by varying the iteration. The proposed model HCEFSO-Trans-Bi-LSTM-GRU has a lower value when assimilated over other conventional techniques that have further enhanced the performance of the given model.

6.4. Validating of the given model for datasets 1 to 3 among various algorithms

The validation of the performance of the given model for datasets 1 to 3 among various algorithms have been depicted in Fig. 5 to 7. In Fig. 5 (a), the value of accuracy by varying the epochs at 250 for the proposed -Trans-Bi-LSTM-GRU model is 7%, 7%, 6%, and 6% higher than that of the WOA-Trans-Bi-LSTM-GRU, HBA-Trans-Bi-LSTM-GRU, CSA-Trans-Bi-LSTM-GRU, and EFO-Trans-Bi-LSTM-GRU models. Thus, it has improved the performance of the efficient IoT-based heart disease prediction model.

6.5. Validating the given model for datasets 1 to 3 among various classifiers

Figure 8 to 10 indicated the validation of the performance of the given model for datasets 1 to 3 among various classifiers. On considering the positive measure F1-Score, the value of the suggested model HCEFSO-Trans-Bi-LSTM-GRU has a higher value over other conventional models by 11%, 10%, 6%, and 7% for BILSTM, Fuzzy-LSTM, CNN-BILSTM, and Trans-BILSTM-GRU model at 50 in dataset 1. It has been similar for all other datasets also. Thus, the prediction rate for the heart disease model has been effectively increased with maximized accuracy.

6.6. Statistical analysis over the given model for dataset 1 to 3

Table 3 has represented the statistical analysis for the given model for datasets 1 to 3 in terms of various algorithms. For dataset 3, the value of the mean for the developed -Trans-Bi-LSTM-GRU model is 0.1%, 0.7%, 0.6%, and 0.4% higher than that of WOA-Bi-LSTM-GRU, HBA-Bi-LSTM-GRU, CSA-Bi-LSTM-GRU, and EFO-Bi-LSTM-GRU models. Hence, it has improved the performance of the newly developed model.

Table 4
Overall validation of the efficient IoT-based heart disease prediction framework in terms of algorithms for datasets 1 to 3.

Measures WOA-Trans-Bi-LSTM-GRU [25] HBA-Trans-Bi-LSTM-GRU [14] CSA-Trans-Bi-LSTM-GRU [7] EFO-Trans-Bi-LSTM-GRU [46] HCEFSO-Trans-Bi-LSTM-GRU

Dataset 1

“Accuracy” 92.49 92.65 93.79 93.43 97.05

“Sensitivity” 92.51 92.65 93.80 93.41 97.06

“Specificity” 92.46 92.65 93.78 93.45 97.04

“Precision” 92.46 92.64 93.77 93.44 97.04

“FPR” 7.54 7.35 6.22 6.55 2.96

“FNR” 7.49 7.35 6.20 6.59 2.94

“NPV” 92.51 92.66 93.81 93.42 97.07

“FDR” 7.54 7.36 6.23 6.56 2.96

“F1-Score” 92.48 92.64 93.79 93.43 97.05

“MCC” 84.97 85.30 87.58 86.86 94.10

Dataset 2

“Accuracy” 92.08 92.74 93.40 93.40 97.36

“Sensitivity” 92.12 92.12 93.33 93.33 96.97

“Specificity” 92.03 93.48 93.48 93.48 97.83

“Precision” 93.25 94.41 94.48 94.48 98.16

“FPR” 7.97 6.52 6.52 6.52 2.17

“FNR” 7.88 7.88 6.67 6.67 3.03

“NPV” 90.71 90.85 92.14 92.14 96.43

“FDR” 6.75 5.59 5.52 5.52 1.84

“F1-Score” 92.68 93.25 93.90 93.90 97.56

“MCC” 0.84 0.85 0.87 0.87 0.95

Dataset 3

“Accuracy” 92.31 92.64 93.65 93.31 96.99

“Sensitivity” 92.71 92.71 93.75 93.75 96.88

“Specificity” 92.12 92.61 93.60 93.10 97.04

“Precision” 84.76 85.58 87.38 86.54 93.94

“FPR” 7.88 7.39 6.40 6.90 2.96

“FNR” 7.29 7.29 6.25 6.25 3.13

“NPV” 96.39 96.41 96.94 96.92 98.50

“FDR” 15.24 14.42 12.62 13.46 6.06

“F1-Score” 88.56 89.00 90.45 90.00 95.38

“MCC” 82.97 83.64 85.82 85.14 93.18

Measures	WOA-Trans-Bi-LSTM-GRU [25]	HBA-Trans-Bi-LSTM-GRU [14]	CSA-Trans-Bi-LSTM-GRU [7]	EFO-Trans-Bi-LSTM-GRU [46]	HCEFSO-Trans-Bi-LSTM-GRU
Dataset 1
“Accuracy”	92.49	92.65	93.79	93.43	97.05
“Sensitivity”	92.51	92.65	93.80	93.41	97.06
“Specificity”	92.46	92.65	93.78	93.45	97.04
“Precision”	92.46	92.64	93.77	93.44	97.04
“FPR”	7.54	7.35	6.22	6.55	2.96
“FNR”	7.49	7.35	6.20	6.59	2.94
“NPV”	92.51	92.66	93.81	93.42	97.07
“FDR”	7.54	7.36	6.23	6.56	2.96
“F1-Score”	92.48	92.64	93.79	93.43	97.05
“MCC”	84.97	85.30	87.58	86.86	94.10
Dataset 2
“Accuracy”	92.08	92.74	93.40	93.40	97.36
“Sensitivity”	92.12	92.12	93.33	93.33	96.97
“Specificity”	92.03	93.48	93.48	93.48	97.83
“Precision”	93.25	94.41	94.48	94.48	98.16
“FPR”	7.97	6.52	6.52	6.52	2.17
“FNR”	7.88	7.88	6.67	6.67	3.03
“NPV”	90.71	90.85	92.14	92.14	96.43
“FDR”	6.75	5.59	5.52	5.52	1.84
“F1-Score”	92.68	93.25	93.90	93.90	97.56
“MCC”	0.84	0.85	0.87	0.87	0.95
Dataset 3
“Accuracy”	92.31	92.64	93.65	93.31	96.99
“Sensitivity”	92.71	92.71	93.75	93.75	96.88
“Specificity”	92.12	92.61	93.60	93.10	97.04
“Precision”	84.76	85.58	87.38	86.54	93.94
“FPR”	7.88	7.39	6.40	6.90	2.96
“FNR”	7.29	7.29	6.25	6.25	3.13
“NPV”	96.39	96.41	96.94	96.92	98.50
“FDR”	15.24	14.42	12.62	13.46	6.06
“F1-Score”	88.56	89.00	90.45	90.00	95.38
“MCC”	82.97	83.64	85.82	85.14	93.18

Table 5

Overall validation for the efficient IoT-based heart disease prediction framework in terms of classifier for datasets 1 to 3.

Measures	BILSTM [24]	FUZZY-LSTM [26]	CNN-BILSTM [24,34]	Trans BILSTM-GRU [24,48]	HCEFSO-Trans-Bi-LSTM-GRU
Dataset 1
“Accuracy”	90.10	92.50	95.11	93.99	97.05
“Sensitivity”	90.10	92.53	95.11	94.01	97.06
“Specificity”	90.09	92.47	95.12	93.98	97.04
“Precision”	90.08	92.47	95.11	93.98	97.04
“FPR”	9.91	7.53	4.88	6.02	2.96
“FNR”	9.90	7.47	4.89	5.99	2.94
“NPV”	90.11	92.54	95.11	94.01	97.07
“FDR”	9.92	7.53	4.89	6.02	2.96
“F1-Score”	90.09	92.50	95.11	93.99	97.05
“MCC”	80.19	85.00	90.23	87.99	94.10
Dataset 2
“Accuracy”	89.77	92.41	95.38	93.73	97.36
“Sensitivity”	89.70	92.73	95.76	93.94	96.97
“Specificity”	89.86	92.03	94.93	93.48	97.83
“Precision”	91.36	93.29	95.76	94.51	98.16
“FPR”	10.14	7.97	5.07	6.52	2.17
“FNR”	10.30	7.27	4.24	6.06	3.03
“NPV”	87.94	91.37	94.93	92.81	96.43
“FDR”	8.64	6.71	4.24	5.49	1.84
“F1-Score”	90.52	93.01	95.76	94.22	97.56
“MCC”	79.43	84.71	90.69	87.37	94.69
Dataset 3
“Accuracy”	89.97	92.64	95.32	93.98	96.99
“Sensitivity”	90.63	92.71	94.79	93.75	96.88
“Specificity”	89.66	92.61	95.57	94.09	97.04
“Precision”	80.56	85.58	91.00	88.24	93.94
“FPR”	10.34	7.39	4.43	5.91	2.96
“FNR”	9.38	7.29	5.21	6.25	3.13
“NPV”	95.29	96.41	97.49	96.95	98.50
“FDR”	19.44	14.42	9.00	11.76	6.06
“F1-Score”	85.29	89.00	92.86	90.91	95.38
“MCC”	78.03	83.64	89.42	86.50	93.18

Table 6

Estimation of the efficient IoT-based heart disease prediction framework using recent existing approaches for all datasets.

Measures	PRCNN [25]	MDLSTM [14]	LWAMCNet [22]	HCEFSO-Trans-Bi-LSTM-GRU
Dataset 1
“Accuracy”	91.29	92.58	94.45	97.05
“Sensitivity”	91.30	92.59	94.46	97.06
“Specificity”	91.28	92.56	94.45	97.04
“Precision”	91.27	92.55	94.44	97.04
“FPR”	8.72	7.44	5.55	2.96
“FNR”	8.70	7.41	5.54	2.94
“NPV”	91.31	92.60	94.46	97.07
“FDR”	8.73	7.45	5.56	2.96
“F1-Score”	91.29	92.57	94.45	97.05
“MCC”	82.58	85.15	88.91	94.10
Dataset 2
“Accuracy”	90.92	92.57	94.39	97.36
“Sensitivity”	90.91	92.42	94.55	96.97
“Specificity”	90.94	92.75	94.20	97.83
“Precision”	92.30	93.85	95.12	98.16
“FPR”	9.06	7.25	5.80	2.17
“FNR”	9.09	7.58	5.45	3.03
“NPV”	89.33	91.11	93.54	96.43
“FDR”	7.70	6.15	4.88	1.84
“F1-Score”	91.60	93.13	94.83	97.56
“MCC”	81.74	85.07	88.70	94.69
Dataset 3
“Accuracy”	91.14	92.64	94.48	96.99
“Sensitivity”	91.67	92.71	94.27	96.88
“Specificity”	90.89	92.61	94.58	97.04
“Precision”	82.66	85.58	89.19	93.94
“FPR”	9.11	7.39	5.42	2.96
“FNR”	8.33	7.29	5.73	3.13
“NPV”	95.84	96.41	97.21	98.50
“FDR”	17.34	14.42	10.81	6.06
“F1-Score”	86.93	89.00	91.65	95.38
“MCC”	80.50	83.64	87.62	93.18

Table 7

Overall analysis of the recommended model using recent methods.

TERMS	SqueezeNet [5]	ACNN-LSTM [11]	DCAlexNet-CNN [36]	HCEFSO-Trans-Bi-LSTM-GRU
Dataset 1
Accuracy	90.53	92.13	94.13	97.05
Sensitivity	90.32	92.06	94.15	97.06
Specificity	90.73	92.20	94.12	97.04
Precision	89.94	91.55	93.63	97.04
FPR	9.27	7.80	5.88	2.96
FNR	9.68	7.94	5.85	2.94
NPV	91.08	92.67	94.60	97.07
FDR	10.06	8.45	6.37	2.96
F1-Score	90.13	91.81	93.89	97.05
MCC	0.81	0.84	0.88	94.10
Dataset 2
Accuracy	86.63	87.70	88.13	97.36
Sensitivity	86.48	87.80	87.93	96.97
Specificity	86.79	87.60	88.35	97.83
Precision	87.11	87.97	88.62	98.16
FPR	13.21	12.40	11.65	2.17
FNR	13.52	12.20	12.07	3.03
NPV	86.15	87.42	87.63	96.43
FDR	12.89	12.03	11.38	1.84
F1-Score	86.80	87.88	88.27	97.56
MCC	0.73	0.75	0.76	94.69
Dataset 3
Accuracy	93.83	95.00	96.40	96.99
Sensitivity	93.78	95.00	96.35	96.88
Specificity	93.89	95.00	96.45	97.04
Precision	93.72	94.87	96.35	93.94
FPR	6.11	5.00	3.55	2.96
FNR	6.22	5.00	3.65	3.13
NPV	93.95	95.13	96.45	98.50
FDR	6.28	5.13	3.65	6.06
F1-Score	93.75	94.93	96.35	95.38
MCC	0.88	0.90	0.93	93.18

6.7. Overall evaluation of the proposed model in terms of algorithms and classifiers

The overall validation for the proposed model by varying the algorithms and classifiers has been depicted in Tables 4 and 5. In dataset 2, for the algorithm, the suggested HCEFSO-Trans-Bi-LSTM-GRU model has 5%, 5%, 3%, and 4% higher in terms of precision when assimilated with other models like WOA-Trans-Bi-LSTM-GRU, HBA-Trans-Bi-LSTM-GRU, CSA-Trans-Bi-LSTM-GRU, and EFO-Trans-Bi-LSTM-GRU. Consequently, on considering the positive measure MCC, the value of the suggested model HCEFSO-Bi-LSTM-GRU has a higher value over other conventional models by 16%, 10%, 4%, and 7% for BILSTM, Fuzzy-LSTM, CNN-BILSTM, and Trans-BILSTM-GRU model at 50 in dataset 1.

6.8. Estimation of the proposed model using recent baseline approaches

An estimation of the offered method using recent approaches is shown in Table 6. In dataset 1, the accuracy of the suggested HCEFSO-Trans-Bi-LSTM-GRU model is secured at 6.31%, 4.83%, and 2.75% elevated performance than PRCNN, MDLSTM, and LWAMCNet. Accordingly, the FNR rate of the designed HCEFSO-Trans-Bi-LSTM-GRU model attains 66.66%, 60.02%, and 44.40% better than PRCNN, MDLSTM, and LWAMCNet for dataset 2. Though, the prediction rate for the heart disease model attained enhanced performance when compared to the other baseline approaches.

6.9. Comparative analysis of the offered model

The overall analysis of the designed model with recent methods are validated for detecting the heart disease is shown in Table 7. The performance shows 7.8%, 5.9%, and 3.6% better performance than SqueezeNet, ACNN-LSTM, and DCAlexNet CNN. Here, the table analysis shows that the developed model is validated with various performance metrics. Moreover, the DCAlexNet CNN model achieves a second better performance. The empirical analysis of the developed model attains better performance than recent methods.

6.10. Discussion of obtained results

The explanation according to the outcomes of the designed method over baseline algorithms and classifiers is depicted as below. From Fig. 6.3, the convergence analysis of the designed method attains enriched performance regarding a number of iterations. Here, the WOA algorithm attains a second better performance, and also, the WOA algorithm attains a lower performance. Due to this, it has the chance to fall into the local optimum. On considering the evaluation of the baseline algorithms, the efficacy of the offered HCEFSO-Trans-Bi-LSTM-GRU- based prediction rate of heart disease based on the IoT model is computed regarding a number of epochs. Here, the EFO algorithm attains a second better performance, and also the HBA algorithm attains a lower performance. However, the HBA algorithm has generated real-world constrained problems. From Fig. 7, while taking the computation of the classifiers the CNN-BiLSTM achieves a second better performance, and also the Bi-LSTM attains a lower performance. It generates overfitting and misclassification issues. The outcomes of the designed method revealed that it is statistically significant. While taking the overall performance of the designed method attains 97.05 % accuracy and 97.03% precision rate.

7. Conclusion

In recent years, heart disease is the prime cause of death in the world. However, the prediction of cardiac disease is a challenging task in the medical field. Medical diagnosis is considered as an important, and also it is quite complex to provide accurate treatment effectively. Especially, a recent technology based on IoT is performed for precisely detecting heart diseases. Here, the IoT collects the sensor information of cardiac disease using recent applications in healthcare organizations. Thus, this paper has effectively implemented the prediction of heart disease based on the IoT. Initially, the data was collected from the standard datasets, and then the preprocessing of raw data was also performed over the aggregated data. Then, the relevant features of heart diseases were chosen with the help of the HCEFSO algorithm, which was integrated by the CSA and EFO. Then, the optimal features were forwarded to the training process, in which the techniques termed as -Trans-Bi-LSTM-GRU were adopted. Here, the weights were updated with the HCEFSO algorithm while validating the training phase. The trained Bi-LSTM-GRU network was utilized in the testing phase for the prediction of heart diseases. In dataset 2, for the algorithm, the suggested HCEFSO-Trans-Bi-LSTM-GRU model has 5%, 5%, 4%, and 4% higher regarding accuracy when assimilated with other models like WOA-Trans-Bi-LSTM-GRU, HBA-Trans-Bi-LSTM-GRU, CSA-Trans-Bi-LSTM-GRU, and EFO-Trans-Bi-LSTM-GRU. The validation outcomes were aided to show the improved performance of the IoT-based heart disease prediction systems. The final validation of the designed model achieves 97% regarding accuracy. While validating with precision, the developed model attains 93%. Moreover, the IoT platform that uses the HCEFSO algorithm helps the doctor to monitor and diagnose the patients in a precise manner. However, in this work, the deep learning strategies require more amount of labeled data. Here, the marking of medical images is a crucial task and time-consuming. It needs intelligent and ensemble techniques to resolve these problems. The synthetic data that are utilized to address the privacy concern as well as to overcome the constraints need to be developed in the future.

References

Al-Makhadmeh

Tolba

, Utilizing IoT wearable medical device for heart disease prediction using higher order Boltzmann model: A classification approach, Measurement. 147 (2019), 106815. doi:10.1016/j.measurement.2019.07.043.

Ali

S.A.

, An Optimally Configured and Improved Deep Belief Network (OCI-DBN) approach for heart disease prediction based on Ruzzo–Tompa and stacked genetic algorithm, IEEE Access 8 (2020), 65947–65958. doi:10.1109/ACCESS.2020.2985646.

Amarbayasgalan

Pham

V.H.

Theera-Umpon

Piao

Ryu

K.H.

, An efficient prediction method for coronary heart disease risk based on two deep neural networks trained on well-ordered training datasets, IEEE Access 9 (2021), 135210–135223. doi:10.1109/ACCESS.2021.3116974.

Balakrishnan

Ambeth Kumar

V.D.

, IoT-enabled classification of echocardiogram images for cardiovascular disease risk prediction with pre-trained recurrent convolutional neural networks, Diagnostics 13(4) (2023), 775.

Balasubramaniam

Kumar

K.S.

Kavitha

Prasanth

Sivakumar

T.A.

, Feature selection and dwarf mongoose optimization enabled deep learning for heart disease detection, Computational Intelligence and Neuroscience (2022).

Basak

Chatterjee

, Smart healthcare surveillance system using IoT and machine learning approaches for heart disease, ASCIS 2022: Advancements in Smart Computing and Information Security 1759 (2023), 304–313.

Braik

M.S.

, Chameleon Swarm Algorithm: A bio-inspired optimizer for solving engineering design problems, Expert Syst. Appl. 174 (2021), 114685. doi:10.1016/j.eswa.2021.114685.

Deperlioglu

Kose

Gupta

Khanna

Sangaiah

A.K.

, Diagnosis of heart diseases by a secure internet of health things system based on autoencoder deep neural network, Comput. Commun. 162 (2020), 31–50. doi:10.1016/j.comcom.2020.08.011.

Dhanamjayulu

Suraj

G.V.

Nikhil

Kaluri

Koppu

, A machine learning algorithm-based IoT-based message alert system for predicting coronary heart disease, ASCIS 2022: Advancements in Smart Computing and Information Security 1759 (2023), 362–376.

10.

Durga

K.P.

Abirami

M.S.

, An IoMT based prediction model for cardiac monitoring using machine learning algorithms, IconDeepCom 2022: Deep Sciences for Computing and Communications 1719 (2023), 103–118. doi:10.1007/978-3-031-27622-4_9.

11.

Dwarakanath

Latha

Annamalai

Kallimani

J.S.

Walia

Belete

, A novel feature selection with hybrid deep learning based heart disease detection and classification in the e-healthcare environment, Computational Intelligence and Neuroscience (2022).

12.

Gong

Zhang

Zhu

Cheng

Zhang

Wang

, Fetal congenital heart disease echocardiogram screening based on DGACNN: Adversarial one-class classification combined with video transfer learning, IEEE Trans. Med. Imaging 39(4) (2019), 1206–1222. doi:10.1109/TMI.2019.2946059.

13.

Guo

Zhang

Liu

Xie

Han

, Recursion Enhanced Random Forest with an Improved Linear Model (RERF-ILM) for heart disease detection on the Internet of medical things platform, IEEE Access 8 (2020), 59247–59256. doi:10.1109/ACCESS.2020.2981159.

14.

Hashim

F.A.

Houssein

E.H.

Hussain

Mabrouk

M.S.

Al-Atabany

, Honey Badger Algorithm: New metaheuristic algorithm for solving optimization problems, Math. Comput. Simul 192 (2022), 84–110. doi:10.1016/j.matcom.2021.08.013.

15.

Islam

M.N.

Raiyan

K.R.

Mitra

Mannan

M.R.

Tasnim

Putul

A.O.

Mandol

A.B.

, Predictis: An IoT and machine learning-based system to predict risk level of cardio-vascular diseases, BMC Health Services Research 23(1) (2023), 171. doi:10.1186/s12913-023-09104-4.

16.

Jadav

N.K.

Gupta

Tanwar

, AI and onion routing-based secure architectural framework for IoT-based critical infrastructure, in: 2023 13th International Conference on Cloud Computing, Data Science & Engineering (Confluence), 2023.

17.

Karhade

Dash

Ghosh

S.K.

Dash

D.K.

Tripathy

R.K.

, Time–frequency-domain deep learning framework for the automated detection of heart valve disorders using PCG signals, IEEE Trans. Instrum. Meas. 71 (2022), 1–11.

18.

Khan

M.A.

, An IoT framework for heart disease prediction based on MDCNN classifier, IEEE Access 8 (2020), 34717–34727. doi:10.1109/ACCESS.2020.2974687.

19.

Kumar

N.A.

Bhatt

B.R.

Anitha

Yadav

A.K.

Devi

K.K.

Joshi

V.C.

, A new diagnosis using a Parkinson’s disease XGBoost and CNN-based classification model using ML techniques, in: 2022 International Conference on Advanced Computing Technologies and Applications, ICACTA, 2022.

20.

Kumar

P.M.

Gandhi

U.D.

, A novel three-tier Internet of Things architecture with machine learning algorithm for early detection of heart diseases, Comput. Electr. Eng. 65 (2018), 222–235. doi:10.1016/j.compeleceng.2017.09.001.

21.

Lavanya

Vidyabharathi

Selva Kumar

Mali

Arunkumar

Aravinth

S.S.

Triny

K.J.

Bhat

J.S.

Tesfayohanis

, Wearable sensor-based edge computing framework for cardiac arrhythmia detection and acute stroke prediction, J. Sens. (2023).

22.

Malibari

A.A.

, An efficient IoT-artificial intelligence-based disease prediction using lightweight CNN in healthcare system, Meas.: Sens. 26 (2023), 100695.

23.

Minh

Q.T.

Phung

P.H.

Huu

P.N.

, Fog-enabled IoT framework for heart disease diagnosis systems, Journal of Mobile Multimedia (2023), 389–418.

24.

Minh-Tuan

Kim

Y.H.

, Bidirectional long short-term memory neural networks for linear sum assignment problems, Applied Sciences 9(17) (2019), 3470.

25.

Mirjalili

Lewis

, The Whale Optimization Algorithm, Adv. Eng. Software 95 (2016), 51–67.

26.

Munagala

N.K.

Langoju

L.R.R.

Rani

A.D.

Reddy

D.R.

, A smart IoT-enabled heart disease monitoring system using meta-heuristic-based fuzzy-LSTM model, Biocybern. Biomed. Eng. 42(4) (2022), 1183–1204. doi:10.1016/j.bbe.2022.10.001.

27.

Ning

Wei

Guo

L.Z.

Chen

, Automatic detection of congestive heart failure based on a hybrid deep learning algorithm in the internet of medical things, IEEE Internet Things J. 8(16) (2021), 12550–12558. doi:10.1109/JIOT.2020.3023105.

28.

Pan

Cheng

Tao

Guo

, Enhanced deep learning assisted convolutional neural network for heart disease prediction on the internet of medical things platform, IEEE Access 8 (2020), 189503–189512. doi:10.1109/ACCESS.2020.3026214.

29.

Pathak

Mandana

Saha

, Ensembled transfer learning and multiple kernel learning for phonocardiogram based atherosclerotic coronary artery disease detection, IEEE J. Biomed. Health. Inf. 26(6) (2022), 2804–2813. doi:10.1109/JBHI.2022.3140277.

30.

Raheja

Manocha

A.K.

, An IoT enabled secured clinical health care framework for diagnosis of heart diseases, Biomed. Signal Process. Control. 80 (2023), 104368. doi:10.1016/j.bspc.2022.104368.

31.

Rajkumar

Devi

T.G.

Srinivasan

, Heart disease prediction using IoT based framework and improved deep learning approach: Medical application, Med. Eng. Phys. 111 (2023), 103937. doi:10.1016/j.medengphy.2022.103937.

32.

Rana

Reddy

Shrivastava

Verma

Ansari

M.S.

Singh

, Secure and smart healthcare system using IoT and deep learning models, in: 2022 2nd International Conference on Technological Advancements in Computational Sciences (ICTACS), 2022, pp. 915–922. doi:10.1109/ICTACS56270.2022.9988676.

33.

Rao

Ramakrishnan

Hassaine

Canoy

Cleland

Lukasiewicz

Salimi-Khorshidi

Rahimi

, An explainable transformer-based deep learning model for the prediction of incident heart failure, IEEE J. Biomed. Health. Inf. 26(7) (2022), 3362–3372. doi:10.1109/JBHI.2022.3148820.

34.

Rhanoui

Mikram

Yousfi

Barzali

, A CNN-BiLSTM model for document-level sentiment analysis, Mach. Learn. Knowl. Extr. 1(3) (2019), 832–847. doi:10.3390/make1030048.

35.

Safa

Pandian

, Applying machine learning algorithm to sensor coupled IoT devices in prediction of cardiac stress – an integrated approach, SRM Institute of Science and Technology (2021), 26.

36.

Saikumar

Rajesh

Sankara Babu

, Heart disease detection based on feature fusion technique with augmented classification using deep learning technology, International Information and Engineering Technology Association 39(1) (2022).

37.

Sarmah

S.S.

, An efficient IoT-based patient monitoring and heart disease prediction system using deep learning modified neural network, IEEE Access 8 (2020), 135784–135797. doi:10.1109/ACCESS.2020.3007561.

38.

Shi

Zhang

Zhuang

Liang

Zhao

Wang

Cai

Shen

, Congestive heart failure detection based on attention mechanism-enabled bi-directional long short-term memory model in the internet of medical things, J. Ind. Inf. Integr. 30 (2022), 100402.

39.

Shuvo

S.B.

Ali

S.N.

Swapnil

S.I.

Al-Rakhami

M.S.

Gumaei

, CardioXNet: A novel lightweight deep learning framework for cardiovascular disease classification using heart sound recordings, IEEE Access 9 (2021), 36955–36967. doi:10.1109/ACCESS.2021.3063129.

40.

Y.S.

Ding

T.J.

Chen

M.Y.

, Deep learning methods in internet of medical things for valvular heart disease screening system, IEEE Internet Things J. 8(23) (2021), 16921–16932. doi:10.1109/JIOT.2021.3053420.

41.

Tuli

Basumatary

Gill

S.S.

Kahani

Arya

R.C.

Wander

G.S.

Buyya

, HealthFog: An ensemble deep learning based smart healthcare system for automatic diagnosis of heart diseases in integrated IoT and fog computing environments, Future Gener. Comput. Syst. 104 (2020), 187–200. doi:10.1016/j.future.2019.10.043.

42.

Velmurugan

A.K.

Padmanaban

Senthil Kumar

A.M.

Azath

Subbiah

, Machine learning IoT based framework for analysing heart disease prediction, AIP Conf. Proc. 2523(1) (2023).

43.

Wang

, A multi-task neural network architecture for renal dysfunction prediction in heart failure patients with electronic health records, IEEE Access 7 (2019), 178392–178400. doi:10.1109/ACCESS.2019.2956859.

44.

Xiao

Jiang

, Heart coronary artery segmentation and disease risk warning based on a deep learning algorithm, IEEE Access 8 (2020), 140108–140121. doi:10.1109/ACCESS.2020.3010800.

45.

Yaqoob

M.M.

Nazir

Khan

M.A.

Qureshi

Al-Rasheed

, Hybrid classifier-based federated learning in health service providers for cardiovascular disease prediction, Appl. Sci 13(3) (2023), 1911.

46.

Yilmaz

Sen

, Electric fish optimization: A new heuristic algorithm inspired by electrolocation, Neural Comput. Appl. (2019).

47.

Zhang

Liu

Xia

Gao

Zhang

de Albuquerque

V.H.C.

, A physics-guided deep learning approach for functional assessment of cardiovascular disease in IoT-based smart health, IEEE Internet Things J. (2023), 1–1.

48.

Zhao

Wang

Yan

Mao

Shen

Wang

, Machine health monitoring using local feature-based gated recurrent unit networks, IEEE Trans. Ind. Electron. 65(2) (2018), 1539–1548. doi:10.1109/TIE.2017.2733438.