Received 2 September 2023, accepted 26 October 2023, date of publication 31 October 2023, date of current version 3 November 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3328909

A Robust Heart Disease Prediction System Using

Hybrid Deep Neural Networks
MANA SALEH AL RESHAN 1 , SAMINA AMIN 2 , MUHAMMAD ALI ZEB 2,

ADEL SULAIMAN 3 , HANI ALSHAHRANI 3 ,

AND ASADULLAH SHAIKH 1 , (Senior Member, IEEE)
1 Department Information Systems, College of Computer Science and Information Systems, Najran University, Najran 61441, Saudi Arabia
2 Institute
of Computing, Kohat University of Science and Technology, Kohat 26000, Pakistan
3 Department Computer Science, College of Computer Science and Information Systems, Najran University, Najran 61441, Saudi Arabia

Corresponding author: Samina Amin ([email protected])

This work was supported by the Deanship of Scientific Research at Najran University through the Research Priorities and Najran Research
Funding Program under Grant NU/NRP/SERC/12/35.

ABSTRACT Heart Disease (HD) is recognized as the leading cause of worldwide mortality by the
World Health Organization (WHO), resulting in the loss of approximately 17.9 million lives each year.
HD prediction is found to be a challenging issue that can provide a computerized estimate of the level
of HD so that additional action can be simplified. Early detection and accurate prediction of HD play a
critical role in providing timely medical interventions and improving patient outcomes. Thus, HD prediction
has expected massive attention worldwide in healthcare environments. Deep Learning (DL) based systems
played a significant role in various disease prediction and diagnosis with good efficiency. To this end, the
main contribution of this paper is to design a robust HD prediction system using Hybrid Deep Neural
Networks (HDNNs) involves combining multiple neural network architectures to extract and learn relevant
features from the input data. The HDNN is employed to apply its feature learning capabilities and non-linear
technology to capture complex patterns and relationships in HD datasets, leading to enhanced prediction
accuracy. For this, three DL models, namely Artificial Neural Networks (ANN), Convolutional Neural
Networks (CNN), Long Short-Term Memory (LSTM), and a new HDNN model combining both CNN and
LSTM along with additional Dense layers are proposed, to develop the hybrid HD prediction architecture.
The proposed models were evaluated on two publicly available HD datasets, including the Cleveland HD
dataset, and a large public HD dataset (Switzerland + Cleveland + Statlog + Hungarian + Long Beach VA).
Additionally, the proposed system was measured through comparison with conventional systems concerning
sensitivity, Matthews Correlation Coefficient (MCC), F1-measure, accuracy, precision, AUC, and specificity.
The promising accuracy achieved through the proposed system is 98.86%. The results demonstrated that this
approach proved more accurate in its predictions than previous research. These outcomes suggest that the
proposed HDNN system has great potential to be embedded into healthcare systems to develop advanced
and reliable HD prediction models that can significantly contribute to medical diagnosis and improve patient
care.

INDEX TERMS Cardiovascular disease, heart disease prediction, Cleveland heart disease dataset, deep
learning, hybrid deep neural networks, CNN-LSTM.

I. INTRODUCTION surge, making it the primary cause of death in most countries

Over the past decade, the prevalence of HD, also known as globally. HD comprises a variety of factors that affect the
Cardiovascular Disease (CVD), has experienced a substantial structure or function of the heart, making it difficult for med-
ical practitioners to make a prompt and accurate diagnosis.
The associate editor coordinating the review of this manuscript and To help doctors identify HD/CVD faster and more effec-
approving it for publication was Mohamed Elhoseny . tively, it is vital to use digital technologies. As a result, the
2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
121574 VOLUME 11, 2023
M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 1. Types of HD (Figure source [2]). FIGURE 2. Cardiovascular risk factors: Generated by biorender.com.

incorporation of computerized technologies in HD diagno- in intelligent healthcare systems is the processing of medical
sis has become crucial to support doctors in making faster images [8], electronic health records [9], gene research [10],
and more accurate assessments. These technological devel- and disease detection in text data [11], which are all carried
opments can support the fight against this widespread and out with DL. For the prediction and diagnosis of diverse dis-
potentially fatal disorder by raising diagnostic accuracy, eases, many expert systems and DL techniques have recently
enabling prompt therapies, and eventually improving patient gained prominence in medical decision support systems [12],
outcomes [1]. [13], [14], [15], [16].
As per the WHO, CVDs/HD is recognized as the lead- Motivated by the recent advancements in DL for medical
ing cause of worldwide mortality, claiming approximately decision support systems [17], [18], this work introduces
17.9 million lives annually. This category encompasses an enhanced and novel Hybrid Deep Neural Networks
various heart and blood vessel disorders, including coro- (HDNNs) system, utilizing a larger and a smaller dataset to
nary HD, cerebrovascular disease, rheumatic HD, and other design the system effectively. By integrating DL approaches
related conditions. HD encompasses various types, including into the existing ML methodologies, researchers can poten-
coronary artery HD, congestive heart failure, heart valve tially achieve more accurate and reliable predictive methods.
HD, cardiomyopathy, heart arrhythmia, and pericarditis as DL technologies can effectively process and analyze large
depicted in Figure 1 [2], and its origins often stem from mul- volumes of medical data, leading to enhanced precision in
tiple factors, including stress, the intake of poor foods, living detecting HD, predicting patient outcomes, and supporting
a sedentary lifestyle, overweight/obesity, lack of exercise, clinical decision-making. The current state of research in
high blood pressure, smoking, diabetes, and binge drinking the field of HD prediction reveals that none of the existing
alcohol, etc. [3], [4], and can be illustrated in Figure 2. studies have applied DL approaches such as deep ANN,
Notably, heart attacks and strokes account for more than LSTM, CNN, and HDNN e.g., CNN with LSTM models
80% of CVD-related fatalities, and a frightening one-third of for HD prediction. A robust HD prediction system using
these deaths occur in those under the age of 70. In order to HDNN offers several advantages over traditional ML meth-
lessen the impact of HD/CVDs on public health, there is an ods. However, it is essential to acknowledge that traditional
urgent need for effective HD prevention, early detection, and ML methods can still be valuable for HD prediction, espe-
improved management. Therefore, early detection of heart cially when data is limited, or interpretability is a crucial
failure is essential since it gives researchers the chance to test factor. In some cases, simpler models like logistic regression
and create efficient pharmaceutical and lifestyle therapies. or decision trees may be preferred, especially when the focus
This is particularly significant due to the fact it may help is on interpretability. Ultimately, the choice between HDNNs
prevent or delay the progression of HD, lowering the risk of and traditional ML models depends on the specific require-
mortality [5]. ments of the HD/CVD prediction task, the availability of data,
A subset of machine learning (ML) known as ‘‘deep learn- computational resources, interpretability needs, and the level
ing’’ (DL) often considers multiple layers of information- of complexity in the relationships within the data.
processing stages in hierarchical structures. DL approaches Thus, developing a robust HD/CVD prediction system
have always been of interest because of their accuracy in using HDNNs involves combining multiple neural network
problem-solving [6], [7]. A good example of the use of DL architectures to extract and learn relevant features from the

VOLUME 11, 2023 121575

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

input data. In this study, a combination of CNN and LSTM II. RELATED WORK
networks, along with additional Dense layers, to create the Researchers have used various data mining approaches,
hybrid architecture will be leveraged. For instance, we use a ML methodologies, and neural networks to develop an HD
combination of a 1D CNN, an LSTM layer, and additional prediction approach [13], [16]. For example, Tarawneh and
Dense layers to create an HDNN for HD/CVD prediction. Embarak [21] designed an HD prediction mechanism based
To operate large-scale medical datasets. As the amount of on the hybrid approach utilizing data mining techniques,
available data grows, the HDNN can be able to amend and and Nalluri et al. [22] presented a chronic HD prediction
operate well without compromising computational efficiency. employing data mining methods, including XGB and LR
For this, we compile and train the model using the two HD algorithms. Gokulnath and Shantharajah [23] designed an
datasets, including the UCI Kaggle Cleveland HD dataset optimized feature selection model using a genetic algorithm
[19], and a comprehensive HD dataset such as (Switzerland and SVM to diagnose HD. Li et al. [24] developed ML based
+ Cleveland + Statlog + Hungarian + Long Beach VA) diagnosis system for HD in e-healthcare. They used ANN,
dataset [20]. After training, the model’s performance will SVM, DT, Logistic Regression (LR), K-Nearest Neighbours
be evaluated by applying various metrics, mainly precision, (KNN), and Naive Bayes (NB) for designing the diagnosing
recall, accuracy, MCC, F1-measure, specificity, ROC curve, system. Compared to the other methodologies, the SVM
and AUC score. method achieved an accuracy of 92.37%. Bharti et al. [25]
The main contributions of the proposed HDNN are defined built the HD prediction system by leveraging ML and DL
as follows: methods. For method building, they used the UCI Cleveland
1. To design a new robust HD prediction model that HD dataset to compare the results and analysis. The DT
achieves high accuracy and diagnoses HD effectively. method accomplished a specificity of 83.1%, and a sensitivity
The HDNN is employed to apply its feature learning of 82.3%. Atallah et al. [26] conducted research work on an
capabilities and non-linear technology to capture com- ensemble method using a majority voting ensemble model
plex patterns and relationships in HD datasets, leading for predicting HD in medical decision systems. The model
to enhanced prediction accuracy. was trained on medical test data of healthy and ill patients
2. To outperform traditional ML methods, namely Sup- recorded in a local clinic. Their ensemble voting achieved an
port Vector Machine (SVM), Decision Tree (DT), accuracy of 90% using a hard voting ensemble technique.
K-Nearest Neighbor (KNN), and Random Forest (RF) These days, data analysis in the healthcare industry may
in HD prediction tasks. By applying deep ANN, LSTM, save lives by enhancing medical diagnostics. Additionally,
CNN, and a hybrid CNN-LSTM, the model will val- several data mining technologies are now accessible to
idate superior predictive performance compared to scientists and can be utilized to carry out studies and exper-
conventional ML systems, particularly when dealing iments owing to the enormous advancements in artificial
with complex medical records. intelligence. For this, Tougui et al. [27] compared min-
3. To exploit the inherent ability of HDNN (CNN-LSTM) ing frameworks, including MATLAB, RapidMiner, Orange,
to automatically learn relevant features from complex Knime, Weka, and Scikit-Learn, applying ML methods,
data. The algorithm can find significant and informa- mainly LR, KNN, SVM, NB, ANN, and RF in order to
tive features that contribute to HD prediction without classify HD. The models were trained on the UCI ML HD
the requirement for manual feature engineering. dataset consisting of 303 instances and 13 features. For the
4. To create an HD prediction system that can be effec- models’ performance, they used three evaluation criteria such
tively integrated into clinical practice. The proposed as accuracy, specificity, and sensitivity. Mahmud et al. [28]
HDNN system can provide accurate and reliable pre- examined various ML models, mainly SVM, KNN, LR, RF,
dictions that can assist healthcare specialists in making DT, and XGB. Researchers used a Kaggle dataset with 70,000
informed decisions, facilitating earlier diagnosis, and different data values for the analysis. The experimental find-
improving patient outcomes. ings illustrated that the RF framework had the best disease
5. To evaluate and compare the performance of the prediction accuracy, with a value of 84.03%. Lutimath et al.
proposed HDNN against existing state-of-the-art HD [29] leveraged SVM to predict the HD from the HD patients
prediction approaches, including conventional ML dataset designed by the UCI ML repository. Pawlovsky [30]
methods. introduced an ensemble method using KNN for diagnosing
The rest of the article is categorized in the following man- HD. The ensemble has been employed with two forma-
ners: Section II discusses a background study conducted on tions. Their ensemble method achieved an average accuracy
recent literature for HD/CVD prediction and classification. of 85%, which was assessed with the HD Cleveland data.
Section III discusses the research gap and Section IV demon- Kavitha et al. [31] designed an ML method for predicting HD
strates the proposed model architecture in detail. Section V by using the UCI HD Cleveland repository and regression
presents experimental results conducted on the two bench- and classification approaches. For model implementation,
mark datasets and their evaluations. Finally, Section VI RF, DT, and a hybrid of RF and DT were employed. The
provides conclusions. hybrid approach obtained an accuracy of 88.7% by predicting

121576 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

HD. Another related work by Shah et al. [32], in which the Hamdaoui et al. [40] established a clinical decision support
authors proposed an ML-based model for HD prediction. For system for predicting HD to help clinicians with diagno-
model training, they used NB, DT, KNN, and RF models by sis and make better decisions. For model building, they
using the UCI HD Cleveland dataset. employed ML algorithms including, NB, KNN, SVM, RF,
Moreover, Almazroi et al. [33] designed a framework and DT for predicting HD. They conducted several experi-
using the DL-based method ANN to predict HD. For model ments to predict HD using the UCI HD dataset, and the result
evaluation, they used four HD datasets, including Cleveland, shows that NB outperforms employing both cross-validation
Hungarian, Switzerland, and Long Beach. For model com- and train-test split methods with an accuracy of 82.17%, and
parison, they used ML techniques and a DL-based ANN 84.28%, respectively. Amin et al. [41] built an HD predic-
model. The experimental results showed that in all datasets tion mechanism based on ML methods that can predict HD
DL achieved the highest accuracy of 83%. This shows a effectively. Prediction models were developed using seven
system for supporting medical decisions is performing at an classification methods, including LR, KNN, NB, DT, SVM,
appropriate level. This performance is a result of the DL NN, and Vote (a hybrid method with NB and LR). After eval-
model’s hidden layers, where the error rate is minimized. uation, their results show that the HD prediction voting model
The suggested DL framework’s greater accuracy suggests attained an accuracy of 87.4%. Bizimana et al. [42] designed
that it may forecast HD effectively. Mohan et al. [34] devel- an ML-based HD prediction approach that makes use of a
oped an ML-based model by finding significant features variety of data scaling techniques, split ratios, ideal param-
for predicting HD. The prediction system was established eters, and ML technologies using UCI HD data. Similarly,
with an ML approach such as hybrid RF with a linear Saboor et al. [43] designed ML based HD prediction system.
approach and achieved an accuracy of 88.7%. Similarly, To validate and compare ML models, they employed Clas-
Chowdhury et al. [35] discussed an ML-based model for sification And Regression Trees (CART), Adaptive Boost
predicting HD by using an HD dataset consistent with Classifier (ABC), LR, ETC, MNB, SVM, RF, LDA, and
564 instances and 18 features. They collected the dataset by Extreme Gradient Boosting (XGB).
personally visiting hospitals and healthcare facilities in the Numerous ML techniques have been used by scholars in
Sylhet region of Bangladesh. Using the benchmark dataset, the field of HD diagnosis and prediction. These conven-
they trained the model by applying ML approaches includ- tional ML techniques have demonstrated notable efficacy
ing, DT, KNN, LR, NB, and SVM. After evaluation and in producing predictive strategies for HD. However, these
comparison among the models, the SVM accomplished bet- prediction models’ accuracy can be further increased with the
ter with an accuracy score of 91%. Rani et al. [36] built introduction of DL and HDNN technology.
a hybrid framework that can detect HD using the clinical
parameters of the heart patient. In their work, a hybridized III. RESEARCH GAP
feature selection method incorporating a genetic approach Early diagnosis plays a crucial role in achieving the goal of
and recursive feature elimination has been employed. For the predicting HD promptly. In the pursuit of diagnosing and
development of the model, SVM, NB, LR, RF, and Adaptive predicting HD/CVD, existing studies have employed ML
boosting techniques were applied. For model training, the techniques. In the current era, HD remains a leading cause
Cleveland HD dataset was used. Among the ML techniques, of mortality globally, posing significant challenges for CVD
the model achieved the best results with the RF with an prediction in clinical data analysis. The continuous growth in
accuracy of 86.6%. the size and complexity of medical datasets in the healthcare
Moreover, Goyal [37] proposed HD prediction methods industry imposes the use of automated systems based on
Lion Optimization-Based Feature Selection (LOFS)-ANN, DL technologies to assist medical authorities/specialists in
LOFS-SVM, and LOFS-DT utilizing the LOFS technique making precise and efficient decisions.
and ML algorithms. The datasets utilized were from the UCI However, one of the major challenges faced by ML
repository. However, Tasnim and Habiba [38] leveraged ML approaches is the decline in accuracy when dealing with
techniques, including NB, SVM, KNN, DT, LR, ANN, and large datasets. To address this issue, this research aims to
RF to estimate the likelihood of coronary HD. Scientists have introduce innovative feature optimization and classification
been working tirelessly to develop an innovative healthcare methodologies for HD prediction by using small datasets
system. A computerized system that can identify HD risk with 303 instances and large datasets with 1190 instances,
could be considered a major accomplishment. They used enabling medical practitioners to achieve early and accurate
the dataset from the UCI ML repository to evaluate their disease diagnosis. This research follows a similar path but
suggested strategy. The feature selection technique enhances introduces an enhanced and innovative HDNN approach,
the performance of traditional ML methods. Among the ML utilizing a larger and a small dataset to train the model
models, RF model with PCA achieved the best accuracy of effectively. The current state of research in the field of HD
92.85% for classifying HD. Kadhim and Radhi [39] pre- prediction reveals that none of the current research studies
sented an ML-based model using RF, SVM, KNN, and DT have applied DL approaches such as deep ANN, LSTM,
algorithms. Compared to the ML models the RF performed CNN, and HDNN e.g., CNN with LSTM models for HD pre-
better than other models with an accuracy score of 94.958%. diction. However, traditional ML and data mining approaches

VOLUME 11, 2023 121577

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 3. Overall flow diagram of the proposed system.

have shown promising results in predicting and classifying HD effectively. To achieve this goal, the proposed approach
HD into HD positive and HD negative. Motivated by the employs deep ANN, LSTM, CNN, and hybrid CNN with
promising development in DL methods, the main objective of LSTM with multiple layers to achieve high analytical and
this study is to design an HDNN system capable of predicting prediction accuracy. To refine HD data and ensure its quality,

121578 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

the proposed HDNN system also considers data imputation The detail of the two datasets is given as follows:
strategies.
To evaluate the performance of the HDNN system, two 1) DATASET-I: CLEVELAND DATASET
benchmark datasets on HD are applied. It is expected that This dataset on HD, which includes 2 classes, 14 attributes,
the HDNN method will greatly increase the accuracy of HD and 303 occurrences, was made available by the Medical
diagnosis when applied to these datasets. Two distinct com- Centre and the Cleveland Clinic Foundation and may be
parisons are presented in the study to demonstrate the efficacy obtained in the UCI repository [19]. A multivariate dataset
of the proposed HDNN. One comparison is carried out with with 76 variables, the Cleveland HD dataset was taken into
the individual method, where the hybrid CNN-LSTM model consideration for this investigation. The Cleveland HD in
is tested against deep ANN, LSTM, and CNN methods. The particular has 76 properties and 302 occurrences. However,
other comparison compares the performance of the proposed only 14 out of 76 traits are used in all reported research. These
system with conventional ML techniques using deep neural 14 characteristics are regarded as the Cleveland database’s
networks. By applying DL approaches, the goal of this work subset. The 13 attributes are all input attributes, and they stand
is to improve HD diagnosis capabilities and forecasting accu- alone. The final column contains an output feature, which
racy. It is expected that the application of data imputation and is essentially a label attribute, and it is based on the input
DL methods will enhance our understanding of HD detection characteristics. For example, if it is 0, the patient is detected
and make early and precise diagnosis easier. The objective is with HD negative and if it is 1 it means that the patient is
to investigate efficient methods to promote clinical diagnosis diagnosed with HD positive.
and early treatment, and to construct an HDNN-based HD
prediction system associated with the satisfaction of health 2) DATASET-II: (SWITZERLAND + CLEVELAND + STATLOG +
specialists. The improved performance of the proposed DL HUNGARIAN + LONG BEACH VA)
methodology over current techniques will be illustrated by a This HD dataset was designed by fusing five well-known HD
comparison with traditional and HDNNs. Furthermore, this datasets that are freely accessible on IEEE Data Port [20].
research has the potential to contribute to the field of HD pre- It may also be found in the Kaggle repository [48]. In this
diction by exploring the capabilities of HDNN techniques and dataset, five HD datasets are integrated over 11 features
improving the accuracy and effectiveness of HD diagnosis which yield it the largest HD dataset accessible so far for
hopefully ultimately advancing clinical diagnosis and early research uses. The dataset is a combination of (Switzerland +
treatment. Cleveland + Statlog + Hungarian + Long Beach VA) dataset
with important features such as maximum heart rate achieved,
IV. PROPOSED HEART DISEASE PREDICTION serum cholesterol, chest pain type, fasting blood sugar, and so
METHODOLOGY on. This dataset on HD comprises 11 features, 1190 instances,
This section demonstrates HD prediction using proposed and 2 classes. It is a collaborative effort involving contribu-
models such as deep ANN, LSTM, CNN, and a combi- tions from the Cleveland Clinical Foundation, the Medical
nation of CNN-LSTM architectures. Figure 3 demonstrates Center, the Long Beach V Clinical Foundation, the Hungarian
the working flow of the proposed HDNN system. The pro- Institute of Cardiology, Switzerland, and the Stalog (Heart)
posed system is implemented using the Anaconda framework Foundation. Table 1 outlines the total number of instances for
(Jupyter Notebook) [44], Python programming language, and each dataset.
fundamental ML/DL libraries, including Scikit-learn [45],
NumPy [46], and TensorFlow [47], etc. Furthermore, a brief TABLE 1. Total number of instances in HD datasets.
demonstration of the suggested HD prediction system is pro-
vided in the following subsections.

A. HEART DISEASE DATASET

The HD datasets used in the suggested HDNN prediction
system are taken from the widely used and freely acces-
sible ML UCI data collection, which has been endorsed
by numerous academics. In this work, two benchmark HD
datasets are utilized, including the Cleveland HD dataset,
and a comprehensive HD dataset such as (Switzerland +
Cleveland + Statlog + Hungarian + Long Beach VA). These B. PREPROCESSING
datasets contain particular characteristics that will be utilized Data preprocessing was done following data collection. The
to decide whether or not to diagnose HD in the patient. The Cleveland dataset had 6 records with missing values. The
output feature’s label attribute has categorical values in the dataset was reduced from 303 to 297 records by eliminating
form of HD positive and HD negative values. HD negative all the records with missing values. Next, the multiclass
values were changed to 0, and HD positive ones were changed values of the predicted attribute for the presence of HD in
to 1. the dataset were converted to binary values (0 for absence;

VOLUME 11, 2023 121579

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

1 for the presence of HD). To execute the data preparation TABLE 2. Feature importance in HD datasets.
operation, all diagnosis values between 2 and 4 were trans-
formed to 1. Thus, the diagnosis values in the resulting dataset
are merely 0 and 1, with 0 denoting the absence of HD and
1 denoting its existence.
Since most medical data is discontinuous, data standardiza-
tion is crucial to converge the data’s features. Data must also
be normalized or standardized before the implementation of
DL approaches. One common procedure for data standardiza-
tion is z-score normalization, which uses the attribute’s mean
(µi ) and standard deviation (σi ) to normalize the attribute’s
data. Data standardization is a method of transforming vari-
ous types of data into a format that is normalized and uniform.
When µi and σi are taken into account as the mean and
standard deviation of the ith attribute of a dataset, the z-score
zij for the jth instance, ij is determined in Equation 1.
xij − µi
zij = (1)
σi E. PREDICTION MODELS
After determining the features, the models were built using
C. FEATURE SELECTION
the four DL prediction and categorization techniques, includ-
The method of feature selection involves picking features
ing ANN, LSTM, CNN, and Hybrid CNN-LSTM. The
from a wide range of available qualities or features in order
proposed models are described in the following sections.
to reduce computational latency and complexity while boost-
ing accuracy. Using the Model Characteristics property, the
1) ARTIFICIAL NEURAL NETWORK
importance of each feature in HD datasets can be determined.
Every function of the outcomes is given a score by feature An ANN [49] is a computational system designed with a
value, demonstrating its relevance and impact on the perfor- large number of simple but well-connected processing units
mance variable. A higher score denotes greater importance or as shown in Figure 4. Variations in external inputs are used
suitability of the feature. To extract the most important fea- to process these units. The ANN architecture proposed in
tures from the dataset, in this study, the Extra Tree Classifier this study was developed by applying multiple weighted
(ETC) is exercised using the Gini relevance technique. This hidden layers directed by a feed-forward network with a back-
model is particularly useful for feature significance analysis propagation algorithm. This approach is commonly used in
as it is equipped with a built-in class that enables feature many applications [50], including image processing, speech
importance assessment, especially in Tree-Based models. recognition, robotic control, sentiment analysis, forecasting,
By using the ETC, we can identify the key features that sig- and power system safety and control management are all
nificantly contribute to the system’s predictive performance areas where ANNs are best suited. ANN can be linked to
and gain valuable insights for our analysis. The significance the human brain. The human brain is made up of several
of various features may have changing values as a result of neurons that function quickly. Each input (bit or signal) of
the randomness of feature samples as can be investigated in
Table 2.

D. DATA SPLITTING
Data splitting is a technique that involves dividing a dataset
into smaller subsets [8] and the normalized preprocessed HD
data is partitioned into two chunks (train, and test set). The
HD dataset is partitioned into an 80:20 holdout validation
technique, with 80% of the statistics being employed to train
the proposed HDNN system and 20% being preserved for
model evaluation (testing). Where, Train set is a statistic
of data samples employed for learning, to train or fit the
parameters of the model (e.g., a real dataset is applied to
train the model). A Test set is a statistic of data samples
that are applied only to evaluate the efficiency of the HDNN
system. FIGURE 4. Structure of ANN.

121580 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 5. Structure of CNN.

data passes via a neuron, where it is interpreted, computed, used to downsample a specified input dimension to reduce
and processed before being passed on to the next neuron the number of variables. One common technique in pooling
cell. While each neuron or node’s overall processing speed is max pooling, which retains the greatest value within a given
is sluggish, the network as a whole is extremely quick and input area. Finally, the Fully Connected layer functions as a
effective. ANNs designed in this study along with the input classifier, enabling data classification based on the informa-
layer, hidden layers, and neurons are listed in table. The first tion obtained from the convolution and max pooling layers.
hidden layer has the same number of nodes as the input layer,
as recommended by the best practices. Five neurons make
3) LONG SHORT-TERM MEMORY
up the second hidden layer, and two neurons make up the
output layer for the prediction of HD cases. The activation LSTM belongs to the domain of DL. It falls under the cat-
function for the hidden layers is a ReLU algorithm, while the egory of recurrent neural networks and is renowned for its
output layer uses a softmax activation function (e.g., binary ability to grasp long-term dependencies, particularly in tasks
classification). involving sequence prediction [11] and [50]. As it progresses,
it takes input and transmits it to others. The LSTM’s cells
perform a variety of tasks. LSTM has a memory state that can
2) CONVOLUTIONAL NEURAL NETWORK remember information and learn long-term dependencies for
A CNN is a DL technology capable of learning directly from long periods. As a result, LSTMs have proven to be a valuable
data. CNNs prove to be exceptionally valuable in recognizing tool in various domains, including speech recognition, time
patterns within images, thereby identifying various objects series forecasting, and natural language processing.
[8], [51]. Moreover, these networks can also be extremely LSTM has a significant advantage over RNN as it incorpo-
advantageous when it comes to categorizing non-image data rates a cell state to store long-term information. This allows
such as audio, time series, and signal data. The primary data from previous time steps to be retained and connected
concept behind CNNs is to extract local attributes from to data in the current time step within the LSTM network.
higher-level data streams and pass them to lower levels to To achieve this, LSTM employs three gates: the input gate, the
detect more complex features. A typical CNN design involves forget gate, and the output gate. The present input is denoted
convolution, pooling, and fully connected layers, as shown in as it , while Ct and Ct−1 represent the current and previous cell
Figure 5. states, respectively. Similarly, Ht and Ht−1 represent the past
In the convolution layer, a collection of kernels is utilized and present outputs.
to generate feature mappings in the form of a tensor. These Figure 6 illustrates the internal architecture of an LSTM,
kernels use ‘‘stride(s)’’ to convolve the information, resulting showcasing how the gates and cell state facilitate memory
in output volumes of integer sizes. However, due to striding, retention and information flow throughout the network. The
the sizes of the convolutional layer’s input volumes decrease. LSTM layer receives the feedback obtained from the dropout
To preserve zero padding is required to fill the input space layer. The calculation is trumped up of four parts: an input
with zeros when the input volume has low-level features. gate (it ), a forget gate (ft ), an output gate (ot ), and a new mem-
The Rectified Linear Unit (ReLU) function introduces ory container (ct ). To integrate the performance of forward
nonlinearity into the feature maps. It calculates the activation and backward, the element-wise computation is calculated
by setting the threshold input to zero. The pooling layer is based on Equations (2 – 5):

VOLUME 11, 2023 121581

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 6. Structure of LSTM.

An LSTM receipts a current input (Xt ), a previous state multivariate HD features were reduced through a max-pooling
(ht−1 ), does some calculation (Equations. 2–5), and afterward layer with 2 × 2 kernels. The resulting feature map was
combines data in the arrangement of a hidden state (h), ⃗ then passed to the LSTM layer in the final phase to extract
as shown below: patient data. The output shape was molded to comprehend the
convolution section (none, 8, 8, 512). The LSTM layer’s input
ft = σ Uf ht−1 , Wf X t + bf


(2) size was lowered by operating a reshaping method (16, 512).
it = σ (Ui ht−1 , Wi X t + bi ) (3) The architecture analyzed the temporal features before pass-
at = tanh (Uc ht−1 , Wc X t + bc ) = tan ât


(4) ing the multivariate HD features via a fully connected layer
ot = σ (Uo ht−1 , Wo X t + bo ) (5) to classify each instance into two categories such as HD
positive, or HD negative (no HD).
4) HYBRID CNN-LSTM NETWORK
V. RESULTS AND DISCUSSION
This study proposed a novel strategy for predicting HD
This section summarizes the experimental findings for all
instances using multivariate HD datasets, including Cleve-
four models’ performance in terms of prediction, optimiza-
land and a comprehensive (Switzerland + Cleveland +
tion, and computational cost for two HD datasets used in med-
Statlog + Hungarian + Long Beach VA). The model involved
ical diagnosis. The efficiency of the mentioned HDNN sys-
combining CNN-LSTM networks, where the CNN was
tem is assessed through the metrics for accuracy, precision,
responsible for capturing intricate audio information, while
sensitivity, MCC, specificity, f-measure, and ROC/AUC.
the LSTM served as the prediction model. To design the
These metrics are determined by equations 5-11 [52]:
CNN-LSTM mechanism with multivariate HD datasets, fea-
tures of HD-related symptoms were extracted from raw the Accuracy
datasets and used for training. TP + TN
The implied hybrid model for HD predicted, as shown = (6)
FN + FP + TP + TNF
in Figure 7, comprised 30 layers: 18 convolutional layers, Recall
12 pooling layers, 1 fully connected layer, 1 LSTM layer,
TP
and an output layer using the softmax approach. Each con- = (7)
volutional block included a pooling layer, two to three 2D FN + TP
CNNs, and one convolutional block, followed by a dropout Preccision
layer with a 20% dropout rate. The features, initiated by TP
= (8)
the ReLU technique, were extracted using a convolutional FP + TP
layer with a 3 × 3 kernel size. The dimensions of the input F − measure

121582 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 7. Hybrid structure of CNN-LSTM.

Recall × Precision
=2× (9) metric may not be sufficient. Therefore, other metrics like
Recall + Precision precision, sensitivity, MCC, specificity, f-measure, ROC, and
Specificity AUC are also calculated to assess the efficiency of the HDNN.
TN Furthermore, Table 4 presents the results of the proposed
= × 10 (10)
FN + TN models when using a larger public HD dataset, which com-
MCC prises data from (Switzerland + Cleveland + Statlog +
TP × TN −FP × FN Hungarian + Long Beach VA). Notably, the DL models,
=√ particularly the HDNN CNN-LSTM, demonstrated superior
(TP+FP) × (TP+FN ) × (TN +FP) × (TN +FN )
(11) performance compared to other models in this comprehensive
dataset. Following the results in Tables 3 and 4, it can be
The experimental analysis of the results obtained with observed that the DL models, especially the HDNN CNN-
different combinations of preprocessing techniques and the LSTM, surpassed the other approaches in terms of various
performance of four ML methods (SVM, KNN, DT, and RF) performance metrics. This underscores the efficiency and
and four DL models (ANN, LSTM, CNN, and CNN-LSTM) effectiveness of the suggested HDNN for HD diagnosis and
are presented in Table 3 and Table 4. prediction, making it a promising approach in the field of
Table 3 reports the scores of key performance metrics medical research and decision-making support.
concerning accuracy, specificity, precision, MCC, sensitivity, Figure 8 illustrates the graphical comparison of the ML
f-measure, and AUC, for different ML and DL methods using algorithms. The comparison indicates that the RF model
the Cleveland dataset. The last row of Table 3 highlights surpassed other models like SVM, KNN, and DT in terms
the best-performing model. However, determining the best of precision, sensitivity, MCC, specificity, and f-measure
model solely based on the highest value of one performance when using the Cleveland dataset. Figure 9 indicates the

VOLUME 11, 2023 121583

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

TABLE 3. Evaluation of the proposed DL models using a comprehensive HD dataset (Switzerland + Cleveland + Statlog + Hungarian + Long Beach VA).

TABLE 4. Evaluation of the proposed DL models using a comprehensive HD dataset (Switzerland + Cleveland + Statlog + Hungarian + Long Beach VA).

FIGURE 8. Graphical comparison of ML models using the Cleveland dataset.

graphical comparison of the proposed DL technologies using For the more extensive dataset (Switzerland + Cleveland +
the Cleveland dataset. The comparison reveals that the Statlog + Hungarian + Long Beach VA), Figure 10 exhibits
CNN-LSTM model achieved superior performance in terms the graphical comparison of ML models. The RF model once
of precision, sensitivity, MCC, specificity, and f-measure again performed better compared to SVM, KNN, and DT
when compared to other models like ANN, LSTM, and CNN. models when considering precision, sensitivity, MCC, speci-

121584 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 9. Graphical comparison of DL models using the Cleveland dataset.

FIGURE 10. Graphical comparison of DL models using a comprehensive dataset. (Switzerland +

Cleveland + Statlog + Hungarian + Long Beach VA.)

ficity, and f-measure. In Figure 11, the graphical comparison The statistical analysis indicates that the results obtained
of the anticipated DL models is presented using the com- from the proposed HDNNs are highly significant compared
prehensive dataset (Switzerland + Cleveland + Statlog + to other techniques. These findings highlight the effectiveness
Hungarian + Long Beach VA). Similar to the Cleveland of the proposed ML and DL models in predicting HD, mak-
dataset, the CNN-LSTM model showed better performance ing them promising approaches for medical decision support
in terms of precision, sensitivity, MCC, specificity, and systems.
f-measure when compared to other models like ANN, LSTM, The results proved that the suggested HDNN system out-
and CNN. performed the state-of-the-art systems. When compared to

VOLUME 11, 2023 121585

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 11. Graphical comparison of DL models using a comprehensive dataset. (Switzerland + Cleveland +
Statlog + Hungarian + Long Beach VA.)

FIGURE 12. ROC/AUC score of ML models using Cleveland.

FIGURE 13. ROC/AUC score of DL models using Cleveland.

traditional ML approaches and existing state-of-the-art sys-

tems, the proposed CNN-LSTM model achieved significantly This suggests that the CNN-LSTM model is highly effective
higher accuracy, precision, sensitivity, MCC, specificity, in predicting HD and surpasses existing methods when using
f-measure, and AUC values of 97.75%, 0.9857, 0.9887, a more diverse and extensive dataset.
0.9660, 0.9787, 0.9718, and 0.9885, respectively, when using Figure 12 displays the AUC scores for ML models like
the Cleveland HD dataset. SVM, KNN, DT, and RF, which are 0.8791, 0.8885, 0.8603,
Additionally, when the comprehensive HD dataset, and 0.9120, respectively. These ML models also exhibit
including data from (Switzerland + Cleveland + Statlog + good performance, but the CNN-LSTM model still outper-
Hungarian + Long Beach VA), was used, the CNN-LSTM forms them with a higher AUC score. On the other hand,
model continued to demonstrate superior performance. in Figure 13, the results demonstrate that the CNN-LSTM
It achieved even higher accuracy, precision, sensitivity, MCC, model outperforms other DL models, including ANN
specificity, f-measure, and AUC values of 98.86%, 0.9913, (95.00%), LSTM (96.00%), and CNN (97.00%), achieving an
0.9874, 0.9705, 0.9942, 0.9983, and 0.9978, respectively. impressive AUC score of 99.00%. The CNN-LSTM strategy

121586 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

performed better in both datasets, indicating that these models

can be beneficial for diagnosing and predicting HD. For both
datasets, the CNN-LSTM system performed exceptionally
well, consistently achieving accuracy values exceeding 95%.
This indicates the HDNN system’s excellence in serving as a
medical decision support system. The high accuracy can be
attributed to the hidden layers of the HDNN system, which
help in reducing the error rate and improving the system’s
overall predictive capabilities. These findings indicate that
the proposed hybrid CNN-LSTM model exhibits superior
performance in both datasets, making it a highly effective
choice for the task at hand. The results from Figure 16
illustrate strong evidence of the capabilities of the proposed
ML and DL methods in enhancing HD/CVD prediction accu-
racy. These results emphasize the significance of leveraging
FIGURE 14. ROC/AUC score of ML models using a comprehensive dataset. advanced HDNNs computational methods, such as CNN-
(Switzerland + Cleveland + Statlog + Hungarian + Long Beach VA.) LSTM, in medical research to improve patient care and
support clinical decision-making in the field of HD/CVD.
Additionally, the performance of the proposed system was
compared with existing studies. The proposed system’s state-
of-the-art comparison was conducted with other existing
works, and the results are listed in Table 5. The outcomes
demonstrated that the hybrid CNN-LSTM achieved a max-
imum accuracy of 98.86% and 97.75%. These simulated
results proved that the CNN-LSTM system outperformed
the existing techniques, showcasing superior performance in
comparison. While most existing works are applied to a single
HD dataset, the proposed system was evaluated on two dif-
ferent HD datasets, each having different ranges and types of
values and features. The outcomes revealed that the suggested
HDNN such as CNN-LSTM consistently demonstrated high
performance across all the datasets. This suggests that the
system is not biased towards any specific dataset or range of
FIGURE 15. ROC/AUC score of DL models using a comprehensive dataset. values and can be employed effectively for various datasets
(Switzerland + Cleveland + Statlog + Hungarian + Long Beach VA.)
for HD prediction at an early stage.
To the best of our knowledge, the accuracy attained by the
CNN-LSTM architecture is the highest among all the state-
shows its superiority in distinguishing between true positive of-the-art systems reported in the literature for both datasets.
and false positive instances in the dataset. This suggests that the anticipated CNN-LSTM mechanism
Furthermore, in Figure 14 and Figure 15, the ROC curves surpasses the performance of other existing algorithms, mak-
illustrate the performance of the proposed ML and DL models ing it a significant advancement in the field of HD prediction.
when using a comprehensive HD dataset that includes data
from (Switzerland + Cleveland + Statlog + Hungarian + A. LIMITATIONS
Long Beach VA). In this case, the hybrid CNN-LSTM model The research work has some limitations, such as its poor
achieves the highest AUC score of 1.00%, surpassing the compatibility with several feature selection techniques, its
AUC scores of ANN, LSTM, and CNN, which are 97.00%, susceptibility to datasets that contain a large amount of miss-
98.00%, and 99.00%, respectively. Figure 14 also presents the ing data, and its lack of comprehensive testing on real-world
ROC scores for ML models such as SVM, KNN, DT, and RF, datasets. Moreover, the lack of deep ensemble learning meth-
which are 0.8844, 0.8815, 0.8331, and 0.8974, respectively. ods in the present model highlights an important path toward
These findings highlight the exceptional accuracy of hybrid improving the research’s predictive power.
CNN-LSTM, ANN, CNN, and LSTM in both datasets, sug-
gesting their potential as useful methods for HD diagnosis B. FUTURE RESEARCH DIRECTIONS
and prediction. There are various methods to improve this study and deal with
A comparison of the suggested ML and DL models for its shortcomings. In the future, we want to further general-
the two datasets is shown in Figure 16. The outcomes ize the system to make it compatible with different feature
show that the ANN, CNN, LSTM, and hybrid CNN-LSTM selection algorithms and more resistant to other datasets with

VOLUME 11, 2023 121587

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

FIGURE 16. Comparison of the proposed ML and DL models for both datasets.

TABLE 5. Comparison of the proposed DL models’ performance against the existing work.

significant amounts of missing data. By carrying out the same approaches for HD/CVD prediction, additional investigation
experiment on an extensive real-world dataset, this study can can be performed. New feature selection techniques can also
be developed further. To evaluate various combinations of DL be used to gain a wider perspective on the important features

121588 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

and boost prediction accuracy. Another approach is to employ [6] F. Ali, S. El-Sappagh, S. M. R. Islam, D. Kwak, A. Ali, M. Imran, and
deep ensemble learning techniques to strengthen and improve K.-S. Kwak, ‘‘A smart healthcare monitoring system for heart disease pre-
diction based on ensemble deep learning and feature fusion,’’ Inf. Fusion,
the model. vol. 63, pp. 208–222, Nov. 2020, doi: 10.1016/j.inffus.2020.06.008.
[7] S. P. Barfungpa, H. K. D. Sarma, and L. Samantaray, ‘‘An intelligent
heart disease prediction system using hybrid deep dense Aquila network,’’
VI. CONCLUSION
Biomed. Signal Process. Control, vol. 84, Jul. 2023, Art. no. 104742, doi:
The heart is crucial for all organisms. The prediction of HD 10.1016/j.bspc.2023.104742.
requires greater accuracy, validity, and reliability because [8] S. Amin, B. Alouffi, M. I. Uddin, and W. Alosaimi, ‘‘Optimizing convo-
even a minor error can result in exhaustion or even death. lutional neural networks with transfer learning for making classification
report in COVID-19 chest X-rays scans,’’ Sci. Program., vol. 2022,
HD-related deaths are common, and the number of these pp. 1–13, May 2022, doi: 10.1155/2022/5145614.
deaths is rising rapidly each year. Early diagnosis plays a [9] S.-C. Huang, A. Pareek, S. Seyyedi, I. Banerjee, and M. P. Lungren,
crucial role in achieving the goal of predicting HD promptly. ‘‘Fusion of medical imaging and electronic health records using deep
learning: A systematic review and implementation guidelines,’’ NPJ Digit.
In the pursuit of diagnosing and predicting HD, existing Med., vol. 3, no. 1, p. 136, Oct. 2020.
studies have employed ML techniques. This research follows [10] B. He, L. Bergenstråhle, L. Stenbeck, A. Abid, A. Andersson, Å. Borg,
a similar path but introduces an enhanced and innovative J. Maaskola, J. Lundeberg, and J. Zou, ‘‘Integrating spatial gene expression
HDNN system, utilizing a larger and a small dataset to train and breast tumour morphology via deep learning,’’ Nature Biomed. Eng.,
vol. 4, no. 8, pp. 827–834, Jun. 2020.
the model effectively. The research showcases the effec- [11] S. Amin, A. Alharbi, M. I. Uddin, and H. Alyami, ‘‘Adapting recurrent
tiveness of the extra tree classifier method, which aids in neural networks for classifying public discourse on COVID-19 symptoms
identifying relevant features for HD prediction. Furthermore, in Twitter content,’’ Soft Comput., vol. 26, no. 20, pp. 11077–11089,
Oct. 2022, doi: 10.1007/s00500-022-07405-0.
the proposed HDNN system employs deep ANN, LSTM, [12] T. Si, J. Bagchi, and P. B. C. Miranda, ‘‘Artificial neural network train-
CNN, and hybrid CNN with LSTM with multiple layers ing using metaheuristics for medical data classification: An experimental
to achieve high diagnosing and prediction accuracy. The study,’’ Expert Syst. Appl., vol. 193, May 2022, Art. no. 116423, doi:
10.1016/j.eswa.2021.116423.
approach also incorporates data imputation techniques to
[13] A. Najafi, A. Nemati, M. Ashrafzadeh, and S. H. Zolfani, ‘‘Multiple-
refine the data and ensure its quality. The proposed system criteria decision making, feature selection, and deep learning: A golden
was trained on two HD datasets such as Cleveland and a triangle for heart disease identification,’’ Eng. Appl. Artif. Intell., vol. 125,
comprehensive HD (Switzerland + Cleveland + Statlog + Oct. 2023, Art. no. 106662, doi: 10.1016/j.engappai.2023.106662.
[14] S. P. Barfungpa, L. Samantaray, H. Kumar Deva Sarma, R. Panda, and
Hungarian + Long Beach VA) dataset. Finally, the discussed A. Abraham, ‘‘D-t-SNE: Predicting heart disease based on hyper param-
system was measured through comparison with conventional eter tuned MLP,’’ Biomed. Signal Process. Control, vol. 86, Sep. 2023,
systems concerning sensitivity, MCC, F1-score, accuracy, Art. no. 105129, doi: 10.1016/j.bspc.2023.105129.
precision, and specificity. The outcomes showed a superior [15] N. V. MahaLakshmi and R. K. Rout, ‘‘Effective heart disease prediction
using improved particle swarm optimization algorithm and ensemble clas-
prediction of the proposed HDNN system compared to the sification technique,’’ Soft Comput., vol. 27, no. 15, pp. 11027–11040,
state-of-the-art. Compared to the traditional ML and existing Aug. 2023, doi: 10.1007/s00500-023-08388-2.
state-of-the-art, the proposed CNN-LSTM achieved the best [16] M. Swathy and K. Saruladha, ‘‘A comparative study of classification and
prediction of cardio-vascular diseases (CVD) using machine learning and
accuracy, precision, sensitivity, MCC, specificity, f-measure, deep learning techniques,’’ ICT Exp., vol. 8, no. 1, pp. 109–116, Mar. 2022,
and AUC of 97.75%, 0.9857, 0.9887, 0.9660, 0.9787, 0.9718, doi: 10.1016/j.icte.2021.08.021.
and 0.9885, respectively, when using Cleveland HD dataset. [17] B. R. Reddy and R. L. Kumar, ‘‘Classification of health care products using
hybrid CNN-LSTM model,’’ Soft Comput., vol. 27, no. 13, pp. 9199–9216,
Furthermore, using a comprehensive HD (Switzerland + Jul. 2023, doi: 10.1007/s00500-023-08279-6.
Cleveland + Statlog + Hungarian + Long Beach VA), the [18] A. Rayan, S. Holyl Alruwaili, A. S. Alaerjan, S. Alanazi, A. I. Taloba,
CNN-LSTM surpass accuracy, precision, sensitivity, MCC, O. R. Shahin, and M. Salem, ‘‘Utilizing CNN-LSTM techniques for
specificity, f-measure, and AUC of 98.86%, 0.9913, 0.9874, the enhancement of medical systems,’’ Alexandria Eng. J., vol. 72,
pp. 323–338, Jun. 2023, doi: 10.1016/j.aej.2023.04.009.
0.9705, 0.9942, 0.9983, and 0.9978, respectively. [19] UCI Machine Learning Repository. Heart Disease Cleveland Daat-
set. Accessed: Jul. 20, 2023. [Online]. Available: https://archive.ics.
uci.edu/dataset/45/heart+disease
REFERENCES
[20] M. Siddhartha. (2020). Heart Disease Dataset (Comprehensive). IEEE
[1] N. C. Long, P. Meesad, and H. Unger, ‘‘A highly accurate firefly based Dataport. Accessed: Jul. 20, 2023. [Online]. Available: https://ieee-
algorithm for heart disease prediction,’’ Expert Syst. Appl., vol. 42, no. 21, dataport.org/open-access/heart-disease-dataset-comprehensive
pp. 8221–8231, Nov. 2015, doi: 10.1016/j.eswa.2015.06.024. [21] M. Tarawneh and O. Embarak, ‘‘Hybrid approach for heart disease predic-
[2] UDMI. Accessed: Jul. 24, 2023. [Online]. Available: https://www.udmi. tion using data mining techniques,’’ in Proc. 7th Int. Conf. Emerg. Internet,
net/cardiovascular-disease-risk/ Data Web Technol. (EIDWT). UAE: Springer, 2019, pp. 447-454.
[3] P. K. Mehta, S. Gaignard, A. Schwartz, and J. E. Manson, ‘‘Tradi- [22] S. Nalluri, R. V. Saraswathi, S. Ramasubbareddy, K. Govinda, and
tional and emerging sex-specific risk factors for cardiovascular dis- E. Swetha, ‘‘Chronic heart disease prediction using data mining tech-
ease in women,’’ Rev. Cardiovascular Med., vol. 23, no. 8, p. 288, niques,’’ in Proc. 3rd Data Engineering and Communication Technology
2022. (ICDECT). Singapore: Springer, 2020, pp. 903–912.
[4] J. M. Alves-silva, M. Zuzarte, and H. Gir, ‘‘The role of essential [23] C. B. Gokulnath and S. P. Shantharajah, ‘‘An optimized feature selection
oils and their main compounds in the management of cardiovas- based on genetic approach and support vector machine for heart dis-
cular disease risk factors,’’ Molecules, vol. 26, no. 3506, pp. 1–26, ease,’’ Cluster Comput., vol. 22, no. S6, pp. 14777–14787, Nov. 2019, doi:
2021. 10.1007/s10586-018-2416-4.
[5] C. Zhou, A. Hou, P. Dai, A. Li, Z. Zhang, Y. Mu, and L. Liu, ‘‘Risk [24] J. P. Li, A. U. Haq, S. U. Din, J. Khan, A. Khan, and A. Saboor,
factor refinement and ensemble deep learning methods on prediction of ‘‘Heart disease identification method using machine learning classification
heart failure using real healthcare records,’’ Inf. Sci., vol. 637, Aug. 2023, in e-healthcare,’’ IEEE Access, vol. 8, pp. 107562–107582, 2020, doi:
Art. no. 118932, doi: 10.1016/j.ins.2023.04.011. 10.1109/ACCESS.2020.3001149.

VOLUME 11, 2023 121589

 M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

[25] R. Bharti, A. Khamparia, M. Shabaz, G. Dhiman, S. Pande, and P. Singh, [45] Scikit-Learn. Accessed: Aug. 4, 2023. [Online]. Available: https://scikit-
‘‘Prediction of heart disease using a combination of machine learning learn.org/stable/
and deep learning,’’ Comput. Intell. Neurosci., vol. 2021, Jan. 2021, [46] NumPy. Accessed: Aug. 4, 2023. [Online]. Available: https://numpy.org/
Art. no. 8387680, doi: 10.1155/2021/8387680. [47] TensorFlow. Accessed: Aug. 4, 2023. [Online]. Available: https://www.
[26] R. Atallah and A. Al-Mousa, ‘‘Heart disease detection using machine tensorflow.org/tutorials/keras/keras_tuner
learning majority voting ensemble method,’’ in Proc. 2nd Int. [48] M. Siddhartha. Heart Disease Dataset (Comprehensive). Kaggle.
Conf. New Trends Comput. Sci. (ICTCS), Oct. 2019, pp. 1–6, doi: Accessed: Aug. 4, 2023. [Online]. Available: https://www.kaggle.com/
10.1109/ICTCS.2019.8923053. datasets/sid321axn/heart-statlog-cleveland-hungary-final
[27] I. Tougui, A. Jilbab, and J. El Mhamdi, ‘‘Heart disease classifica- [49] S. Amin, M. I. Uddin, M. A. Zeb, A. A. Alarood, M. Mahmoud, and
tion using data mining tools and machine learning techniques,’’ Health M. H. Alkinani, ‘‘Detecting information on the spread of dengue on Twitter
Technol., vol. 10, no. 5, pp. 1137–1144, Sep. 2020, doi: 10.1007/ using artificial neural networks,’’ Comput., Mater. Continua, vol. 67, no. 1,
s12553-020-00438-1. pp. 1317–1332, 2021, doi: 10.32604/cmc.2021.014733.
[28] T. Mahmud, A. Barua, M. Begum, E. Chakma, S. Das, and [50] S. Amin, M. I. Uddin, S. Hassan, A. Khan, N. Nasser, A. Alharbi, and
N. Sharmen, ‘‘An improved framework for reliable cardiovascular H. Alyami, ‘‘Recurrent neural networks with TF-IDF embedding technique
disease prediction using hybrid ensemble learning,’’ in Proc. Int. Conf. for detection and classification in tweets of dengue disease,’’ IEEE Access,
Electr., Comput. Commun. Eng. (ECCE), Feb. 2023, pp. 1–6, doi: vol. 8, pp. 131522–131533, 2020, doi: 10.1109/ACCESS.2020.3009058.
10.1109/ECCE57851.2023.10101564. [51] L. Alzubaidi, J. Zhang, A. J. Humaidi, A. Al-Dujaili, Y. Duan,
[29] N. M. Lutimath, C. Mouli, B. K. B. Gowda, and K. Sunitha, ‘‘Pre- O. Al-Shamma, J. Santamaría, M. A. Fadhel, M. Al-Amidie, and L. Farhan,
diction of heart disease using hybrid machine learning technique,’’ in ‘‘Review of deep learning: Concepts, CNN architectures, challenges, appli-
Paradigms of Smart and Intelligent Communication, 5G and Beyond, cations, future directions,’’ J. Big Data, vol. 8, no. 1, p. 53, Mar. 2021, doi:
A. Rai, D. K. Singh, A. Sehgal, and K. Cengiz, Eds. Singapore: Springer, 10.1186/s40537-021-00444-8.
2023, pp. 277–293. [52] P. K. Shrivastava, M. Sharma, P. Sharma, and A. Kumar, ‘‘HCBiLSTM:
[30] A. P. Pawlovsky, ‘‘An ensemble based on distances for a kNN method A hybrid model for predicting heart disease using CNN and BiLSTM
for heart disease diagnosis,’’ in Proc. Int. Conf. Electron., Inf., Commun. algorithms,’’ Measurement, Sensors, vol. 25, Feb. 2023, Art. no. 100657,
(ICEIC), Jan. 2018, pp. 1–4, doi: 10.23919/ELINFOCOM.2018.8330570. doi: 10.1016/j.measen.2022.100657.
[31] M. Kavitha, G. Gnaneswar, R. Dinesh, Y. R. Sai, and R. S. Suraj, ‘‘Heart [53] A. Tiwari, A. Chugh, and A. Sharma, ‘‘Ensemble framework for cardio-
disease prediction using hybrid machine learning model,’’ in Proc. 6th Int. vascular disease prediction,’’ Comput. Biol. Med., vol. 146, Jul. 2022,
Conf. Inventive Comput. Technol. (ICICT), Jan. 2021, pp. 1329–1333, doi: Art. no. 105624, doi: 10.1016/j.compbiomed.2022.105624.
10.1109/ICICT50816.2021.9358597. [54] V. Chang, V. R. Bhavani, A. Q. Xu, and M. Hossain, ‘‘An artificial
[32] D. Shah, S. Patel, and S. K. Bharti, ‘‘Heart disease prediction using intelligence model for heart disease detection using machine learning
machine learning techniques,’’ Social Netw. Comput. Sci., vol. 1, no. 6, algorithms,’’ Healthcare Anal., vol. 2, Nov. 2022, Art. no. 100016, doi:
p. 345, Nov. 2020, doi: 10.1007/s42979-020-00365-y. 10.1016/j.health.2022.100016.
[33] A. A. Almazroi, E. A. Aldhahri, S. Bashir, and S. Ashfaq, ‘‘A clinical
decision support system for heart disease prediction using deep learning,’’
IEEE Access, vol. 11, pp. 61646–61659, 2023.
[34] S. Mohan, C. Thirumalai, and G. Srivastava, ‘‘Effective heart disease
prediction using hybrid machine learning techniques,’’ IEEE Access, vol. 7,
pp. 81542–81554, 2019.
[35] M. N. R. Chowdhury, E. Ahmed, M. A. D. Siddik, and A. U. Zaman, ‘‘Heart
disease prognosis using machine learning classification techniques,’’ in
MANA SALEH AL RESHAN received the B.S.
Proc. 6th Int. Conf. Converg. Technol. (I2CT), Apr. 2021, pp. 1–6, doi:
degree in information systems from King Khalid
10.1109/I2CT51068.2021.9418181.
[36] P. Rani, R. Kumar, N. M. O. S. Ahmed, and A. Jain, ‘‘A decision sup-
University, Abha, Saudi Arabia, in 2007, the M.S.
port system for heart disease prediction based upon machine learning,’’ degree (Hons.) in computer, information, and net-
J. Reliable Intell. Environments, vol. 7, no. 3, pp. 263–275, Sep. 2021, doi: work security from DePaul University, Chicago,
10.1007/s40860-021-00133-6. IL, USA, in 2011, and the Ph.D. degree in com-
[37] S. Goyal, ‘‘Predicting the heart disease using machine learning tech- puter science from The Catholic University of
niques,’’ in Proc. ICT Anal. Appl. (ICT4SD). Singapore: Springer, 2023, America (CUA), Washington, DC, USA, in 2019.
pp. 191–199. He was a Teaching Assistant with the College
[38] F. Tasnim and S. U. Habiba, ‘‘A comparative study on heart disease of Computer Science and Information Systems,
prediction using data mining techniques and feature selection,’’ in Proc. Najran University, Saudi Arabia, from 2007 to 2009. He was a Lecturer
2nd Int. Conf. Robot., Electr. Signal Process. Techn. (ICREST), Jan. 2021, with the College of Computer Science and Information Systems, Najran
pp. 338–341, doi: 10.1109/ICREST51555.2021.9331158. University, in 2012, where he is currently an Assistant Professor and the
[39] M. A. Kadhim and A. M. Radhi, ‘‘Heart disease classification using Head of the Network Engineering Department. His current research interests
optimized machine learning algorithms,’’ Iraqi J. Comput. Sci. Math., include computer network and security, system security, wireless and mobile
vol. 2023, pp. 31–42, Feb. 2023. security, body area networks, and cloud security.
[40] H. E. Hamdaoui, S. Boujraf, N. E. H. Chaoui, and M. Maaroufi, ‘‘A clinical
support system for prediction of heart disease using machine learning
techniques,’’ in Proc. 5th Int. Conf. Adv. Technol. Signal Image Process.
(ATSIP), Sep. 2020, pp. 1–5, doi: 10.1109/ATSIP49331.2020.9231760.
[41] M. S. Amin, Y. K. Chiam, and K. D. Varathan, ‘‘Identification of
significant features and data mining techniques in predicting heart
disease,’’ Telematics Informat., vol. 36, pp. 82–93, Mar. 2019, doi:
10.1016/j.tele.2018.11.007.
SAMINA AMIN received the M.S. degree in computer science from the
[42] P. C. Bizimana, Z. Zhang, M. Asim, and A. A. A. El-Latif, ‘‘An
Institute of Computing, Kohat University of Science and Technology, Kohat,
effective machine learning-based model for an early heart disease pre-
Pakistan, in 2020, where she is currently pursuing the Ph.D. degree. She
diction,’’ Biomed Res. Int., vol. 2023, Jan. 2023, Art. no. 3531420, doi:
10.1155/2023/3531420. has published several articles in reputed journals. Her research interests
[43] A. Saboor, M. Usman, S. Ali, A. Samad, M. F. Abrar, and N. Ullah, include machine learning, deep learning, data science, data mining, natural
‘‘A method for improving prediction of human heart disease using language processing, social media analysis, neural networks, recurrent neural
machine learning algorithms,’’ Mobile Inf. Syst., vol. 2022, Jan. 2022, networks, convolutional neural networks, image processing, sentiment anal-
Art. no. 1410169, doi: 10.1155/2022/1410169. ysis, health informatics, recommender systems, and reinforcement learning.
[44] Anaconda. Accessed: Aug. 4, 2023. [Online]. Available: https:// She has received the Gold Medal from the Kohat University of Science and
anaconda.org/ Technology. She serves as a reviewer for different journals.

121590 VOLUME 11, 2023

M. S. Al Reshan et al.: Robust HD Prediction System Using HDNNs

MUHAMMAD ALI ZEB is currently pursuing HANI ALSHAHRANI received the bachelor’s
the Ph.D. degree with the Institute of Comput- degree in computer science from King Khalid Uni-
ing, Kohat University of Science and Technology, versity, Abha, Saudi Arabia, the master’s degree in
Kohat, Pakistan. He is a Lecturer with the Insti- computer science from California Lutheran Uni-
tute of Computing, Kohat University of Science versity, Thousand Oaks, CA, USA, and the Ph.D.
and Technology. He has published several papers degree from Oakland University, Rochester, MI,
in reputed journals and conferences. His research USA. He is currently an Associate Professor of
interests include machine learning, deep learning, computer science and information systems with
social media analysis, natural language process- Najran University, Najran, Saudi Arabia. His cur-
ing, sentiment analysis, data science, visualization, rent research interests include smartphones, the
topic modeling, neural networks, image processing, and temporal graphs IoT, crowdsourcing security, and privacy.
neural networks.
ASADULLAH SHAIKH (Senior Member, IEEE)
received the B.Sc. degree in software development
from the University of Huddersfield, U.K., the
M.Sc. degree in software engineering and manage-
ment from the University of Gothenborg, Sweden,
and the Ph.D. degree in software engineering from
the University of Southern Denmark. He was a
Researcher with UOC, Barcelona, Spain. He is
currently a Professor, the Head of research and
graduate studies, and the Coordinator of seminars
ADEL SULAIMAN received the master’s degree and training with the College of Computer Science and Information Sys-
in software engineering from the University of tems, Najran University, Najran, Saudi Arabia. He has vast experience in
St. Thomas, USA, and the Ph.D. degree in com- teaching and research. He has more than 170 publications in the area of
puter science from Iowa State University, USA. software engineering in international journals and conferences. His current
He is currently an Assistant Professor with the research interests include UML model verification, UML class diagrams
College of Computer Science and Information verification with OCL constraints for complex models, formal verification,
Systems, Najran University. He is also the Vice and feedback technique for unsatisfiable UML/OCL class diagrams. He is
Dean for quality and development in the deanship also an Editor of the International Journal of Advanced Computer Systems
of e-learning and leads the Research Unit, Col- and Software Engineering (IJACSSE) and an International Advisory Board
lege of Computer Science, Najran University. His of several conferences and journals. Further details can be obtained using
research interests include human–computer interaction, large-scale simula- www.asadshaikh.com.
tions, mixed reality, scientific visualization, and data science.

VOLUME 11, 2023 121591