diagnostics

Article
An Improved Long Short-Term Memory Algorithm for
Cardiovascular Disease Prediction
T.K. Revathi 1 , Sathiyabhama Balasubramaniam 1, *, Vidhushavarshini Sureshkumar 2
and Seshathiri Dhanasekaran 3, *

1 Department of Computer Science and Engineering, Sona College of Technology, Salem 636005, India;
revathi.tk@sonatech.ac.in
2 Department of Computer Science and Engineering, Faculty of Engineering and Technology, SRM Institute
of Science and Technology, Vadapalani Campus, Chennai 600026, India; vidhushs@srmist.edu.in
3 Department of Computer Science, UiT The Arctic University of Norway, 9037 Tromsø, Norway
* Correspondence: sathiyabhama@sonatech.ac.in (S.B.); seshathiri.dhanasekaran@uit.no (S.D.)

Abstract: Cardiovascular diseases, prevalent as leading health concerns, demand early diagnosis
for effective risk prevention. Despite numerous diagnostic models, challenges persist in network
configuration and performance degradation, impacting model accuracy. In response, this paper
introduces the Optimally Configured and Improved Long Short-Term Memory (OCI-LSTM) model
as a robust solution. Leveraging the Salp Swarm Algorithm, irrelevant features are systematically
eliminated, and the Genetic Algorithm is employed to optimize the LSTM’s network configuration.
Validation metrics, including the accuracy, sensitivity, specificity, and F1 score, affirm the model’s
efficacy. Comparative analysis with a Deep Neural Network and Deep Belief Network establishes the
OCI-LSTM’s superiority, showcasing a notable accuracy increase of 97.11%. These advancements
position the OCI-LSTM as a promising model for accurate and efficient early diagnosis of cardiovas-
cular diseases. Future research could explore real-world implementation and further refinement for
seamless integration into clinical practice.

Keywords: cardiovascular disease; long short-term memory; salp swarm algorithm; genetic algorithm;
disease prediction model
Citation: Revathi, T.K.;
Balasubramaniam, S.; Sureshkumar,
V.; Dhanasekaran, S. An Improved
Long Short-Term Memory Algorithm for 1. Introduction
Cardiovascular Disease Prediction. In recent times, a surge in fatalities has been linked to cardiovascular disease, with
Diagnostics 2024, 14, 239. https:// the predominant factor being challenges to the heart’s ability to efficiently pump blood
doi.org/10.3390/diagnostics14030239 throughout the body, causing disruptions in blood circulation [1]. Among the spectrum of
Academic Editor: Fabiano Bini heart-related ailments, cardiovascular disease (CVD) stands out as the most detrimental to
human health. Its escalating prevalence has positioned CVD as a leading cause of height-
Received: 21 November 2023 ened mortality rates, presenting substantial challenges to global healthcare industries [2].
Revised: 17 January 2024
According to surveys, CVD has accounted for a staggering 4 in 10 fatalities, affecting nearly
Accepted: 21 January 2024
17.9 million individuals, with a particularly pronounced impact in Asia [3,4].
Published: 23 January 2024
Various attributes or features, such as sex, age, fasting blood sugar, chest pain [5], chest
pain type, chest pain location, blood sugar level, cigarette habit, depression level, electro-
cardiogram [6], exercise-induced angina (exang), resting electrocardiographic results, slope,
Copyright: © 2024 by the authors.
old peak, heart status, poor diet, cholesterol, obesity, family history, alcohol intake, high
Licensee MDPI, Basel, Switzerland.
blood pressure, and physical inactivity [7–13], have been used in different research studies.
This article is an open access article CVD prediction traditionally relies on invasive methods, depending on a patient’s
distributed under the terms and medical history and an analysis report from a medical scientist. Moreover, it is a challenging
conditions of the Creative Commons and costly process. Non-invasive methods, such as clinical decision-making support models
Attribution (CC BY) license (https:// using Machine Learning (ML) and Deep Learning (DL) approaches, are instrumental in
creativecommons.org/licenses/by/ addressing these issues.
4.0/).

Diagnostics 2024, 14, 239. https://doi.org/10.3390/diagnostics14030239 https://www.mdpi.com/journal/diagnostics

Diagnostics 2024, 14, 239 2 of 19

The combination of CNN–LSTM (Convolutional Neural Network–Long Short-Term

Memory) methods was used to automatically detect COVID-19. This combination employs
three types of X-ray images for disease prediction, with LSTM serving as a classifier to
distinguish various COVID-19 cases [14]. It ensures better results for image datasets of
various sizes and resolutions [15], effectively addressing the issue of overfitting [16,17].
Deep Learning has made significant contributions in various domains, including med-
ical imaging, disease tracking, protein structure analysis, drug discovery, and assessing
virus severity and infectivity to control the COVID-19 outbreak [18,19]. Modern technolo-
gies such as Deep Learning, Machine Learning, and Data Science are contributing to the
fight against all types of deadly diseases [20].
The primary objective of the proposed OCI-LSTM model is to resolve issues that have
been detrimental to the performance of prediction models. The main issue is related to
data training, which leads to overfitting and underfitting. Another issue is optimizing the
network model’s configuration. The model tends to overfit by learning even from small
details in the training data, leading to inadequate results when applied to test data [21,22].
Poor learning on the part of the model results in underfitting, where both training and
testing data produce poor results. The core reasons for these issues lie in the inappropriate
design of the network model and its configuration, as well as in the presence of irrelevant
features. These issues increase both the computational cost and the prediction time for
CVD. To address this, the Salp Swarm Algorithm (SSA) is employed to remove noisy
or duplicate features, helping to find the optimal features effectively. Furthermore, an
improved LSTM is proposed for classification, with the Genetic Algorithm (GA) used to
optimize the network configuration. The GA fine-tunes the model by selecting the right
time window size, offering an optimal solution and enhancing model performance.
Finally, experiments are conducted, and four performance metrics are considered
for model evaluation. The Cleveland dataset from the online UCI repository is used for
training and testing, a dataset commonly employed in heart disease research.

2. Related Works
In this section, we discuss the usage of various optimization algorithms, classifiers,
performance metrics applied, and the results obtained in different research works. Finally,
we identified the gaps observed in the related works.
Latha and Jeeva [23] utilized an ensemble approach, combining multiple classifiers
and employing bagging and boosting techniques to enhance the accuracy of their predic-
tion model. Tao et al. [24] applied Machine Learning techniques to classify ECG signal
recordings, achieving a high accuracy of 94.03%. However, they acknowledged a general-
ization issue in their work, indicating a need for improvement in extending their model’s
applicability beyond the experimental setting.
Arabasadi et al. [25] proposed a hybrid approach integrating the GA with classifiers
for predicting coronary arterial disease. Issues such as the suboptimal learning rate and
momentum factors contributed to this limitation. Pérez et al. [26] introduced latent Dirichlet
allocation for discovering insights from the dataset, while their model was evaluated using
qualitative and quantitative measures.
Chatzakis et al. [27] presented a cardiovascular prediction approach based on ECG
images and patients’ medical history data, forming a Decision Support System (DSS).
However, their focus on maintaining health records limited information about predicting
the CVD risk factors. Mohan et al. [28] proposed a hybridized linear-based Machine
Learning approach using Random Forest to enhance the accuracy. They achieved 88.4%
accuracy on the Cleveland UCI repository but faced challenges due to the absence of
restrictions on feature selection.
Ali et al. [29] employed a Deep Belief Network (DBN) for heart disease prediction, opti-
mizing it with the Ruzzo–Tompa feature selection algorithm. Despite achieving an accuracy
of 94.61%, time complexity issues arose due to suboptimal feature selection. Wang et al. [30]
introduced a Deep Neural Network to address the data imbalance, utilizing a Recurrent
Diagnostics 2024, 14, 239 3 of 19

Neural Network (RNN). While achieving accuracies ranging from 83.84% to 87.54% on
different databases, the model did not determine the optimal size of the time window for
hidden layers.
Mirjalili et al. [31] proposed optimization techniques, the SSA and Multi-Objective
Salp Swarm Algorithm (MSSA), with the MSSA showing high network convergence.
Hsiao et al. [32] utilized a Deep Learning framework for cardiovascular risk prediction,
employing autoencoders for feature selection and softmax for classification. However,
potential network generalization problems were noted.
Abdeldjouad et al. [33] introduced hybridized approaches, including the MOEFC,
Logistic Regression, and AdaBoost, with feature selection using the Wrapper method. The
model did not outperform other Machine Learning models. Gers and Schmidhuber [34]
proposed an LSTM variant, while Chung et al. [35] introduced the Gated Recurrent Unit
(GRU). Altan et al. [36] applied the Hilbert–Huang transform for ECG analysis, but they
did not use feature selection.
Hochreiter and Schmidhuber [37] introduced the LSTM as an RNN with long-term mem-
ory but faced challenges in terms of the computation volume and time costs. Modifications by
healthcare researchers aimed to enhance the LSTM’s performance. Javeed, A. et al. [38] pro-
posed the FWAFE method for feature selection, using ANN and DNN frameworks for heart
disease diagnosis. However, the achieved accuracies ranged widely from 50.00% to 91.83%.
Javeed, A. et al. [39] developed a Machine Learning-based diagnostic system for
coronary artery disease detection. They conducted a systematic review of heart disease
prediction methods but did not propose new work, focusing on comparing previous
methods. Al Bataineh, A. and Manacek, S. [40] developed and compared Machine Learning-
based systems for heart disease prediction using the Cleveland Heart Disease dataset. Their
alternative MLP training technique and PSO algorithm achieved an accuracy of 84.61%.
Hassan, C.A. et al. [41] explored Machine Learning techniques for coronary heart
disease prediction, using 11 classifiers. Random Forest outperformed the others with
a 96% accuracy level. Kurian, N.S. et al. [42] conducted a comparative analysis of Machine
Learning classifiers for heart disease prediction with minimal attributes. They evaluated
Nearest Neighbor, Gradient Boosting, Support Vector Machine, Naive Bayes, Logistic
Regression, and Random Forest, identifying attribute correlation and effectiveness.
Rana, M. et al. [43] employed common Machine Learning methods for heart disease
prediction, using the Kaggle dataset. They provided a comparative analysis of the SVM,
Naïve Bayes, Random Forest, Decision tree, and K-Nearest Neighbor, emphasizing their
utility in classification tasks. Islam, M. et al. [44] presented five supervised Machine
Learning techniques for the Wisconsin Breast Cancer dataset, with ANNs achieving the
highest accuracy, precision, and F1 score.
Hasan, M.K. et al. [45] developed a mathematical model for breast cancer detection
using symbolic regression. They achieved successful detection with minimal errors using
the UCI Machine Learning repository dataset. Ayon, S.I. and Islam, M.M. [46] developed a
Deep Neural Network model for diabetes diagnosis using the PID dataset, demonstrating
high accuracy and performance through cross-validation.
Haque, M.R. et al. [47] presented an expert scheme for liver disorder classification
using RFs and ANNs. They achieved accuracy rates of 80% and 85.29% for RFs and ANNs,
respectively. Ayon, S.I. et al. [48] compared computational intelligence techniques for
coronary artery heart disease prediction, finding that DNN achieved the highest accuracy
of 98.15%.
This literature review exposes the limitations of existing methodologies for predicting
heart disease risk factors, highlighting challenges in effectively mitigating overfitting and
underfitting, employing time-intensive optimization techniques, and relying on traditional
diagnostic tools such as ECG. These research gaps are summarized in Table 1. To address
these challenges, this paper introduces the OCI-LSTM as a solution to prevent cardiovas-
cular disease (CVD). The integration of the LSTM with the GA is intended to enhance
the predictive capabilities. The OCI-LSTM is applied to the well-established Cleveland
Diagnostics 2024, 14, 239 4 of 19

Heart Disease dataset, addressing overfitting through optimal feature selection. The model
also tackles network configuration challenges by randomly determining the number of
suitable layers and hyperparameters. The OCI-LSTM is specifically designed to overcome
the identified issues and elevate the overall model performance.

Table 1. Research gaps.

Author Classification Technique Gap Accuracy (%)

Naïve Bayes (NB), C 4.5, Bayes Net, The decline in accuracy is attributed to
Latha and Jeeva [23] Multilayer Perceptron (MLP), Random the absence of an appropriate feature 85.48
Forest and PART selection algorithm.
XG Boost, K-Nearest Neighbor and The generalization issue remains in this
Tao et al. [24] 94.03
Support Vector Machine (SVM) research work.
The learning rate and momentum
Arabasadi et al. [25] Roulette Wheel method factors have not been optimized to the 78
desired level.
The model’s overall performance has
been compromised due to the
Pérez et al. [26] Decision Support System 91.6
constrained search space for
dimensionality reduction.
Diagnosing CVD is challenging because
the authors have developed only a
Decision Support System (DSS) for
Chatzakis et al. [27] Decision Support System maintaining health records and have 92.3
not provided sufficient details about
their prediction and the classification of
cardiovascular risk factors.
The model exhibits no constraints on
Mohan et al. [28] SVM with an Apriori algorithm feature selection, leading to a 88.4
classification error of up to 11.6%.
Time complexity issues arise in the
Ali et al. [29] Deep Belief Network model due to inadequate 94.61
feature selection.
The attained accuracies, ranging from
Javeed, A. et al. [38] FWAFE-ANN and FWAFE-DNN 50.00% to 91.83%, are notably on the 50.00–91.83
lower side.
Al Bataineh, A. The accuracy obtained reaches a
Multilayer Perceptron 84.61
and Manacek, S. [40] maximum of 84.61%.

3. Proposed OCI-LSTM Approach

We present a new model aimed at improving the accuracy of CVD prediction while
addressing network generalization problems, specifically overfitting and underfitting. Ad-
ditionally, it deals with configuration- and optimization-related issues, such as determining
the optimal network configuration. The process begins with preprocessing using the
min–max scaling algorithm. Then, the best attributes are selected through the SSA. Subse-
quently, these optimized features are fed into the OCI-LSTM, which effectively resolves the
mentioned issues and enhances the prediction accuracy level.

3.1. Min–Max Scaling for Feature Normalization

In this phase, the preprocessing involves normalizing missing values and irrelevant
data using the min–max scaling method. The dataset encompasses various attributes, such
as age, patient gender, type of chest pain, resting and fasting blood pressure, maximum
heart rate, slope of the ST segment, exercise-induced angina, number of primary vessels, etc.
Regular and systematic monitoring aids in comprehensive data collection for the repository.
Diagnostics 2024, 14, 239 5 of 19

However, challenges arise due to missing data values, patient interruptions, and technical
faults during the collection process, impacting disease analysis [49–51].
The collected data, containing missing values and irrelevant information, pose a
potential hindrance to the model’s performance. Therefore, it is crucial to address these
issues by applying the standard scaling method to the entire dataset and evaluating the
distributed data outcomes [52–54]. Following this process, noisy and unnecessary data are
eliminated, retaining only pertinent information.
Upon examination, it is discovered that among the 330 instances in the dataset,
6 instances contain missing values. To ensure standardized results for easy interpreta-
tion during model training, the normalization process is applied, and the standardized
outcome is calculated using Equation (1).

Xvalue − mean
Standardised outcome = × 100 (1)
sd
where Xvalue is the maximum of the heart data value, and Standard Deviation (SD) is the
standard deviation.
All the six instances are handled effectively with the use of Equation (1). The dataset
mean is calculated using Equation (2).

∑in=1 x
mean = (2)
n
After calculating the mean, the SD is computed using the Equation (3).
s
1 n
N i∑
Standard Deviation = ( xi − mean)2 (3)
=1

where N is the total number of samples considered for calculating the SD. Given that the
Cleveland dataset contains features with different ranges and magnitudes, we normalize
the entire dataset, particularly the nominal features. However, the categorical features are
not suitable for the scaling process. Therefore, the min–max approach is used to adjust the
values to a range of 0–1. This adjustment aids the model in interpreting the data easily
during the training phase.
The data normalization is performed as follows:

Xvalue − Xmin _value

N′ = (4)
Xmax _value − Xmin _value

In Equation (4), N ′ is denoted as normalized data, Xvalue is noted as a particular data

of any instance, Xmin _value is represented as the minimum value of the whole dataset, and
Xmax_value is represented as the maximum value of the whole dataset.
This paper extends its analysis by estimating various parameters, such as the variance,
minimum, maximum, correlation, and energy, to mitigate the risk factors associated with
CVD during disease prediction. In the initial step, features that seem to provide no value are
substituted with new ones. The dataset often contains extensive patient information, and
while some features are relevant for disease prediction, others may be deemed irrelevant,
potentially leading to overfitting. To address this, the paper incorporates feature set
reduction along with optimization to enhance the disease recognition process. The Salp
Swarm approach is employed to obtain the most essential optimized features from the
original dataset, as detailed in the following section.

3.2. Salp Swarm for Finding an Optimal Subset Feature

In this section, the focus is on selecting the most appropriate and useful features
for the model to predict disease with greater efficiency. The SSA is employed for the
Diagnostics 2024, 14, 239 6 of 19

purpose of attribute selection, enhancing the model’s learning process by eliminating

unwanted attributes.
The SSA leverages the swarming mechanism observed in salps, a type of marine
organism, to randomly select a population. In the sea, a salp chain, known as a salp swarm,
is formed, with the leader salp positioned at the front end and the remaining salps as
swarm followers. Salp positions are denoted in an n-dimensional search location, where
‘n’ represents the total count of identifiers in a given problem. The feature optimization
process encompasses three steps: 1. initializing the population, 2. updating the leader’s
position, and 3. updating the follower’s position. These steps reflect the clustering process
of a salp swarm. The subsections below provide a detailed discussion of the working
principle of the SSA.

3.2.1. Initializing Population

The population initialization is carried out in the S × D Euclidean workspace, where
S is the swarm scale and D is the dimension in space. Consider the available food in the
space to be fd and it is assigned as fd = [ f d1 , f d2 , . . . , f dd ]T where the position of each salp
can be denoted as Pn = [ Pn,1 , Pn,2 , . . . , Pn,D ] T , where n = 1, 2, . . . , N. The upper and lower
bound is denoted as Ub , Lb . The upper bound is said as Ub = [Ub1 , Ub2 , . . . , UbD ] T and the
lower bound is represented as Lb = [ Lb1 , Lb2 , . . . , LbD ] T .
The random initialization of the population is computed using Equation (5).

XS× D = rand(S × D ) × (Ub − Lb ) + Lb × ones(S × D ) (5)

The leader and followers state of the population in the dth dimension are x1,d and xk,d
where, k = 2, 3, . . . , N.

3.2.2. Updating Leader Position

In a salp swarm, the leader is responsible for finding food in the space. It must also
guide the entire group to move in search of food. It is essential to update the leader’s
position, which is achieved using Equation (6).

x1,d = f dd + r1 ((Ubd − Lbd )r2 + Lbd ) (6)

where r1 and r2 are random numbers within the interval range [0,1]. The leader’s movement,
searching ability, and individual population diversity are randomly enhanced by the
parameters mentioned in Equation (6). In all meta-heuristic approaches, there is a key
parameter known as r1 , as defined in Equation (6). This parameter is also referred to as
the convergence factor. During the iteration process, this parameter balances the trade-off
between exploitation and exploration. If r1 is greater than 1, the algorithm performs global
exploration. If r1 is less than 1, it focuses on local exploration to find an accurate estimation
value. The value of r1 should fall within the range of 2 to 0 for the initial iteration of the
algorithm to conduct global search and subsequently improve the accuracy of the following
iterations. The convergence factor is calculated using Equation (7).

4i 2
r1 = 2e−( imax ) (7)

where i represents the current iteration and imax denotes the total number of iterations.

3.2.3. Updating Follower Position

In the SSA, the followers adopt a series of chain movements rather than random
movements. To determine the followers’ movement, certain important aspects need to
be considered, including the followers’ initial position, speed of motion, and acceleration.
Diagnostics 2024, 14, 239 7 of 19

Newton’s law of motion is followed to calculate the motion distance, and it is computed
using Equation (8).
1
Motion Distance = αi2 + s0 i (8)
2
where i is the iteration during the optimization process, i = 1 when there is discrepancies
happens between iterations, s0 is the speed of the followers, and it becomes 0 at the
first iteration, and α is the followers’ acceleration, as calculated between the first and
last iterations.
The followers’ acceleration is calculated using Equation (9). Always the followers
follow the predecessor salp.  
s f inal − s0
α= (9)
t
So, the salp’s movement speed can be determined using Equation (10).
 
s f inal = xki −1,d − xk,d
i
/t (10)

where t = 1; s0 = 0; hence, the Motion Distance is assigned as Equation (11).

1 i i

Motion Distance = xk−1,d − xk,d (11)
2
The follower position is updated with the help of Equation (12).

i +1 i 1 i 
xk,d = xk,d + Motion Distance = xk,d + xki −1,d (12)
2
i is the dth dimensional kth follower in the ith iteration and x i +1 is representing
where xk,d k,d
the followers’ position in the (i + 1) th iteration. Algorithm 1 describes the flow of the SSA.

Algorithm 1. Salp Swarm Algorithm

1. Initialization:
Salp swarm random population generation Xi where i = 1,2,3, . . ., n
2. Determine each salp’s fitness value.
3. Assign X * as one of the best searching agents.
4. While the end condition has not arrived
5. Update the convergence factor r1 by Equation (7)
6. For each and every salp
7. If (n = = 1)
8. Update leader salp’s position, x1,d using Equation (6)
9. Else
i +1
10. Update the follower salps’ position, xk,d using Equation (12)
11. End if
12. End for
13. Estimate each salp’s fitness value using Equation (14).
14. Update the X * with its finest solution.
15. End while
16. Return X * along with its best fitness value.

3.3. Genetic Algorithm for Optimization

The GA draws inspiration from natural evolution and is categorized as a meta-heuristic
approach [55]. In essence, the GA employs fundamental principles of genetics and evolu-
tion, incorporating crossover and mutation. Optimal solutions in the GA are derived by
selecting the fittest individuals from each generation. The core process of the GA involves
various operators to choose qualified members of the current generation [56,57].
Diagnostics 2024, 14, 239 8 of 19

The selection operator facilitates individuals’ involvement in determining the next

generation based on their fitness. It shapes the subsequent population level by evaluating
the compatibility of the current generation. Stochastic Universal Sampling (SUS) and the
Roulette Wheel (RW) stand out as commonly used selection operators in the GA [58]. The
RW calculates the selection probability for everyone separately. Equation (13) computes
the proportionate fitness selection of an individual:

( f itness)i
( Probability)i = N
(13)
∑k=1 ( f itness)k

where ( f itness)i is denoted as the fitness of an individual i, ( Probability)i is the probability

of an individual, and N represents total individuals involved in the population.
Substitution Operator: This operator facilitates the transfer of members from one
generation to the next and plays a crucial role in the propagation process:

( f itness)i = w1 × Age + w2 × restecg + w3 × maxhr . . . . (14)

where w1, w2, w3 . . . are weights assigned to different health indicators.

Recombination Operator: The recombination operator substitutes substrings of two dif-
ferent members from the same generation using the concept of intersection. Common
approaches for the recombination operator include single-point, two-point, and uniform
crossovers [59].
Mutation Operator: Responsible for changing the genes of members of the current
generation to create the next generation. Various methods are available for performing
mutation, including uniform, non-uniform, boundary, and Gaussian mutations [60]. The
Gaussian operator is commonly used among these methods, adding random values to the
selected gene from a normal distribution. Consider if x ∈ [u, v], a chosen gene is used for
performing the mutation process. Then, x ′ is calculated as given in Equation (15).

x ′ = min(max ( N ( x, φ), u), v) (15)

where N ( x, φ) represents the mutation operation, and φ mutation rate depends on the
time interval.

3.4. OCI-LSTM Model

We propose the OCI-LSTM model, integrating it with the Genetic Algorithm (GA)
to select the optimized time window for the LSTM units. The LSTM offers a significant
advantage by enhancing the model’s performance through the utilization of information
from past events to determine the suitable time window. The selection of an appropriate
window size is crucial; if too small, the network may overlook essential information, and
if too large, the model may become overfitted with training data. Figure 1 illustrates the
GA-based network configuration process.
The OCI-LSTM model consists of two phases. In the first phase, parameters are appro-
priately set. The network includes an input layer and two hidden layers. The GA ensures
that the hidden layers contain the optimal number of hidden neurons. Two activation
functions are implemented in the OCI-LSTM network: the sigmoid function in the input
and hidden nodes to scale input values to the range of −1 to 1, and the linear function
for the output nodes, given the problem’s nature in predicting CVD. Initially, the network
weights receive random values, later adjusted using the Adam optimizer, known for its
computational efficiency [61]. The evolutionary-based search algorithm, the GA, is em-
ployed to determine the optimal window size and explore the architectural factors of the
OCI-LSTM network.
Diagnostics 2024,
Diagnostics 2024,14,
14,x239
FOR PEER REVIEW 9 of 19
9 of 19

Figure 1. Genetic Algorithm-based network configuration process.

The OCI-LSTM model consists of two phases. In the first phase, parameters are ap-
propriately set. The network includes an input layer and two hidden layers. The GA en-
sures that the hidden layers contain the optimal number of hidden neurons. Two activa-
tion functions are implemented in the OCI-LSTM network: the sigmoid function in the
input and hidden nodes to scale input values to the range of −1 to 1, and the linear function
for the output nodes, given the problem’s nature in predicting CVD. Initially, the network
weights receive random values, later adjusted using the Adam optimizer, known for its
computational efficiency [61]. The evolutionary-based search algorithm, the GA, is em-
ployed to determine the optimal window size and explore the architectural factors of the
OCI-LSTM
Figure network.
Figure1.1.Genetic
GeneticAlgorithm-based
Algorithm-based network configuration
network process.
configuration process.
Figure 2 illustrates the workflow of the LSTM. In the second phase of the network
model, Figure
the
The
2 illustrates
evaluation the GA
ofmodel
the workflow
fitnessofisthe LSTM. In Different
conducted. the secondLSTM phaseunits
of theare
network
utilized in
model, theOCI-LSTM
evaluation of the GAconsists
fitness of two phases.Different
is conducted. In the first
LSTM phase,
units parameters
are utilized are ap
thepropriately
hidden layers, andThe various window sizes are applied to the OCI-LSTM for this evalu-
in the hiddenset.
layers, network
and includes
various window ansizes
input
arelayer and
applied totwo
thehidden
OCI-LSTMlayers.
for The
this GA en
ation. The populations,
sures that the
evaluation. The hidden
initially
layers
populations,
assigned
contain
initially
arbitrary
the optimal
assigned
values,
number
arbitrary
undergo
values,of
an initialization
hiddenanneurons.
undergo Twopro-
initialization activa
cess before exploring
tion functions
process the two-dimensional
are implemented
before exploring space
in the OCI-LSTM
the two-dimensional using the operators.
space usingnetwork: the sigmoid function in the
the operators.
input and hidden nodes to scale input values to the range of −1 to 1, and the linear function
for the output nodes, given the problem’s nature in predicting CVD. Initially, the network
weights receive random values, later adjusted using the Adam optimizer, known for its
computational efficiency [61]. The evolutionary-based search algorithm, the GA, is em
ployed to determine the optimal window size and explore the architectural factors of the
OCI-LSTM network.
Figure 2 illustrates the workflow of the LSTM. In the second phase of the network
model, the evaluation of the GA fitness is conducted. Different LSTM units are utilized in
the hidden layers, and various window sizes are applied to the OCI-LSTM for this evalu
ation. The populations, initially assigned arbitrary values, undergo an initialization pro
cess before exploring the two-dimensional space using the operators.

Figure 2. LSTM
Figure workflow
2. LSTM workflowdiagram.
diagram.

TheThe chromosomesininthis
chromosomes thiswork
workareareencoded
encodedasasbinary
binarybits,
bits,representing
representing the
the time
time win-
window’s
dow’s size andsizethe
and the LSTM
LSTM cell counts.
cell counts. Genetic Genetic operators
operators then then
searchsearch forbest
for the the best
solution,
solution, evaluating these solutions using a fitness method. The Mean Squared
evaluating these solutions using a fitness method. The Mean Squared Error (MSE) is em- Error (MSE)
is employed as the fitness function in this work. The smallest MSE value returned by the
ployed as the fitness function in this work. The smallest MSE value returned by the archi-
architectural factors is considered the optimal solution. If the termination condition is
tectural factors
satisfied, is considered
the derived thesolution
optimal optimal orsolution. If the
the nearest termination
optimal solutioncondition
is applied is
tosatisfied,
the
OCI-LSTM model. If not, the entire genetic operations are repeated until the condition is
met. Parameters such as the mutation, crossover, and population size are adjusted during
experimentation to enhance the model’s fitness and improve the results. In this experiment,
the OCI-LSTM uses a crossover parameter value of 0.7 and a mutation rate of 0.15, running
for a total of 10 generations to meet the termination criteria.
Figure 2. LSTM workflow diagram.

The chromosomes in this work are encoded as binary bits, representing the time win
dow’s size and the LSTM cell counts. Genetic operators then search for the best solution
evaluating these solutions using a fitness method. The Mean Squared Error (MSE) is em
ployed as the fitness function in this work. The smallest MSE value returned by the archi
 the
the derived
derived optimal
optimal solution
solution or
or the
the nearest
nearest optimal
optimal solution
solution isis applied
applied toto the
the OCI-LSTM
OCI-LSTM
model.
model. If If not,
not, the
the entire
entire genetic
genetic operations
operations are
are repeated
repeated until
until the
the condition
condition isis met.
met. Param
Param
eters such as the mutation, crossover, and population size are adjusted during
eters such as the mutation, crossover, and population size are adjusted during experimen experimen
tation
tation to
to enhance
enhance thethe model’s
model’s fitness
fitness and
and improve
improve thethe results.
results. In
In this
this experiment,
experiment, the the OCI
OCI
Diagnostics 2024, 14, 239 10 of 19
LSTM
LSTM uses a crossover parameter value of 0.7 and a mutation rate of 0.15, running for a
uses a crossover parameter value of 0.7 and a mutation rate of 0.15, running for
total
total of
of 10
10 generations
generations to to meet
meet the
the termination
termination criteria.
criteria.
Figure
Figure
3 illustrates
Figure3 illustrates
3 illustrates the
the the
proposed
proposed
proposed
OCI-LSTM
OCI-LSTM
OCI-LSTM
using
using using
the Salp
Salp Swarm
theSwarm
the Salp Swarm and
and GA,andwhile
GA,
GA, whil
whil
Figure
Figure 4
4 provides
provides an
an overview
overview of
of the
the entire
entire
Figure 4 provides an overview of the entire framework. framework.
framework.

Figure
Figure3.3.
Figure Proposed
3.Proposed
Proposed OCI-LSTM
OCI-LSTM
OCI-LSTM using
using
using the
the Salp
the Salp Salp Swarm
SwarmSwarm and
and GA.
and GA. GA.

Figure
Figure 4.Framework
Figure4.4. Frameworkdiagram.
Framework diagram.
diagram.

We utilize the publicly available Cleveland dataset from the online UCI repositories [62].
We
We utilize
utilize the
the publicly
publicly available
available Cleveland
Cleveland dataset
dataset from
from the
the online
online UCI
UCI repositorie
repositorie
The dataset comprises a total of 76 attributes or features. However, only 13 out of the
[62]. The dataset comprises a total of 76 attributes or features. However, only
76 attributes are commonly used by most researchers for diagnosing heart disease. Theout of th
[62]. The dataset comprises a total of 76 attributes or features. However, only 13
13 out of th
76 attributes
76 attributes
dataset are commonly
are303
consists of commonly used
used
instances, with by most researchers
by mosthaving
6 instances researchers for
missingfor diagnosing
diagnosing
values. heart
heart disease. Th
disease.
In the preprocessing Th
dataset
dataset consists
consists of
of 303
303 instances,
instances, with
with 6
6 instances
instances having
having missing
missing
phase of the proposed OCI-LSTM process, we begin by removing the outliers from the values.
values. In
In the
the prepro
prepro
cessing
cessingAs
dataset. phase
phase of
of the
the
a result, weproposed
eliminate OCI-LSTM
proposed OCI-LSTM process,
uswe
process,
6 instances, leaving we begin
with 297 by
begin by removing
removing
instances the outliers
outliers from
the with.
to work from
Table 2 provides a neat sketch of 13 features from the Cleveland dataset.
Diagnostics 2024, 14, 239 11 of 19

Table 2. Brief sketch of 13 features from the Cleveland dataset.

S. No Sign Name Data Type Description Range

Age of subject
1 f_1 Age numeric Between 29 to 77
in years
1—male
2 f_2 Sex Binary Gender of subject
0—female
1—typical angina
2—atypical angina
3 f_3 cptype nominal Chest pain type
3—non-anginal pain
4—asymptomatic
Resting blood
4 f_4 restbp numeric [94:200]
pressure
5 f_5 Ser_chol numeric Serum cholesterol [126:564]
Fasting blood 0-false
6 f_6 fastbp Binary
sugar 1-true
0—normal
1—Abnormal ST-T wave
Resting electrocar-
7 f_7 restecg nominal 2— likely/exactly to
diographic
have left
ventricular hypertrophy
Maximum heart
8 f_8 maxhr numeric [71:202]
rate
Exercise-induced 0—no
9 f_9 exerir Binary
angina 1—yes
10 f_10 Op numeric ST depression [0:6.2]
1—upslope
Slope of ST
11 f_11 slopeST nominal 2—flat
segment
3—downslope
Number of major
12 f_12 numvesl nominal 0 to 3
vessels
3—normal
Thalassemia or
13 f_13 Thal nominal 6—fixed defect
defect type
7—reversable defect

The process of selecting the most relevant and necessary features using the SSA is
crucial for enhancing the model’s performance. The seven selected attributes (f_1, f_7, f_8,
f_9, f_10, f_12, and f_13) are considered optimal features based on the SSA. Each of these
attributes plays a significant role in predicting CVD in the Cleveland Heart Disease dataset.
The clinical significance of all these attributes is provided below
f_1 (Age): Age is a well-established risk factor for CVD. The likelihood of devel-
oping cardiovascular issues often increases with age, making it a crucial attribute for
prediction models.
f_7 (Serum Cholesterol): Elevated cholesterol levels are associated with an increased risk
of heart disease. Monitoring serum cholesterol levels helps in assessing cardiovascular health.
f_8 (Fasting Blood Sugar): High fasting blood sugar levels can indicate diabetes, which
is a risk factor for CVD. Monitoring blood sugar is essential for predicting and managing
cardiovascular risk.
f_9 (Resting Electrocardiographic Results): Resting electrocardiographic results pro-
vide insights into the heart’s electrical activity. Certain patterns or abnormalities in ECG
can indicate potential heart issues.
Diagnostics 2024, 14, 239 12 of 19

f_10 (Maximum Heart Rate): The maximum heart rate during exercise is a valuable
indicator. Abnormalities or deviations from the expected maximum heart rate can signal
cardiovascular problems.
f_12 (Exercise-Induced Angina): The presence of angina (chest pain) during exercise
is a significant symptom of coronary artery disease. This attribute helps in identifying
individuals with exercise-related cardiovascular issues.
f_13 (Old Peak): The ‘Old Peak’ attribute refers to the depression induced by exercise
relative to rest. This can be indicative of stress on the heart during physical activity,
providing valuable information for CVD prediction.
These selected attributes collectively provide essential information about the patient’s
age, biochemical profile, cardiac function at rest and during exercise, and presence of
symptoms, which are crucial factors for predicting the risk of cardiovascular disease.
The SSA, by optimizing the feature selection process, ensures that the chosen attributes
contribute significantly to the model’s predictive accuracy while mitigating overfitting and
enhancing overall performance.
The temporal aspects of the data using the proposed network comprise the information
about multiple visits of each patient, the time interval between every visit, the target
variable’s presence or absence of heart disease, and the number of steps that (visits or
intervals) the LSTM network should consider when making predictions.
The class label in the dataset is a multi-class variable that ranges from 0 to 4. Here,
the value 0 represents the absence of CVD, while the values from 1 to 4 represent various
stages of CVD presence. In this study, we follow the approach outlined in [63] to convert
the multi-class label into a binary class label, where label 0 indicates the absence of CVD
and label 1 denotes the presence of CVD.
Upon applying this transformation, it is determined that out of the 297 instances in
the dataset, 164 correspond to healthy subjects/patients, as indicated by label 0 (absence
of CVD).
The cost function of a GA–SSA algorithm is represented in the Equation (16).

cost( x ) = w1 × f GA ( x ) + w2 × f SSA ( x ) (16)

where x represents the solution, which may include genetic parameters and positions of
salps, f GA ( x ) is the objective function or cost associated with the Genetic Algorithm compo-
nent, f SSA ( x ) is the objective function or cost associated with the Salp Swarm Algorithm
component, and w1 and w2 are weights that control the contribution of each algorithm
to the overall cost. The weights w1 and w2 could be used to balance the influence of the
Genetic Algorithm and Salp Swarm Algorithm component on the optimization process.
The minimization of the objective function could be:

f GA ( x ) = Minimize( x )

f SSA ( x ) = Minimize( x )

4. Results and Discussion

This work makes several key contributions:
(i) Introduction of the OCI-LSTM: A novel approach, the OCI-LSTM, is proposed to en-
hance CVD prediction by effectively mitigating both overfitting and underfitting. The
method involves the selection of a pertinent feature subset from various optimization
algorithms, with the Salp Swarm Algorithm demonstrating superior performance in
addressing generalization issues.
(ii) Network Configuration Resolution: The OCI-LSTM model resolves network configu-
ration challenges by identifying temporal patterns, such as the optimal time window
size, and determining the finite LSTM units using the GA. The integration of a local
search with the GA enhances the model iteratively, ensuring the GA discovers the
optimal hidden layers for the LSTM, resulting in the finest OCI-LSTM design.
Diagnostics 2024, 14, 239 13 of 19

(iii) Comparative Analysis: The OCI-LSTM is rigorously analyzed by comparing it with

conventional models, Deep Neural Network (DNN), and Deep Belief Network (DBN).
The results showcase the highest accuracy rate and optimal convergence when com-
pared to the models.
(iv) Outstanding Results with Limited Data: Remarkably, the proposed OCI-LSTM achieves
outstanding results using a small volume of the Cleveland Heart Disease dataset and
a minimal set of parameters. This highlights the model’s efficiency and effective-
ness in CVD prediction, emphasizing its potential for practical implementation in
real-world scenarios.
The dataset underwent two-way partitioning into training and testing sets. The train-
ing set facilitated model training and the testing set was used to evaluate the classifier’s
performance [64,65]. Each partition served as a testing fold to ensure comprehensive
testing and training over 10 folds. The performance evaluation utilized key metrics, in-
cluding the accuracy, sensitivity, precision, and F1 score. These metrics were calculated
using the formulas provided in Equations (17)–(20), respectively [66–69]. The results, av-
eraged over the 10 folds, contributed to a comprehensive assessment of the OCI-LSTM
model’s performance.

Accuracy = ((TN+ TP)) ÷ ((TN+ TP+ FN + FP)) (17)

Sensitivity = TP ÷ (TP + FN) (18)

Precision = TP ÷ (TP + FP) (19)
F1 − score = (2 × precision × sensitivity) ÷ (precision + sensitivity) (20)
where TP is the True Positives, TN is the True Negatives, FP is the False Positives, and FN
is the False Negatives.
The experiment commences by partitioning the entire set of instances in the Cleveland
dataset into two subsets. Out of the total 297 instances, 207 instances are allocated to the
training set, while the remaining 90 instances are designated for testing the model. The
architecture details of the DBN and DNN models are illustrated in Table 3. The distribution
of instances for both the training and testing sets is detailed in Table 4. Table 5 is presented,
which provides a comparative analysis of the results of all three models.

Table 3. Architecture details of DBN and DNN models.

Activation No. of Trainable

Model Layer Type Units Output Size
Function Parameters
Input - (None,13) - -
DBN Hidden Layer 1 50 - Sigmoid 700
Output 1 (None,1) Sigmoid 51
Input - (None,13) - -
Hidden Layer 1 100 - ReLU 200
DNN
Hidden Layer 2 50 - ReLU 220
Output 1 (None,1) Sigmoid 31

Table 4. Summary of the training and testing dataset.

S. No Training Instances Testing Instances Total Instances

1 207 90 297
Diagnostics 2024, 14, 239 14 of 19

Table 5. Overall performance of the models.

Metrics DNN (%) DBN (%) OCI-LSTM (%)

Sensitivity 91.95 96.03 98.78
Specificity 90.54 92.65 95.37
F1 Score 90.73 94.77 97.32
Accuracy 91.72 95.73 97.11

Diagnostics 2024, 14, x FOR PEER REVIEWThe comparative analysis presented in Table 5 indicates a variation in the perfor-
14 of 19
mance of the OCI-LSTM and the other two network models. The results obtained for all
four metrics of the OCI-LSTM are superior when compared to the DNN and DBN with
optimal features.
Table 4. Summary of the training and testing dataset.
From Table 6, it is evident that using the SSA, the classification accuracy of the OCI-
LSTMS.has NoimprovedTraining
when compared
Instances with other famous
Testing feature optimization
Instances algorithms.
Total Instances
Figure 51illustrates the comparison
207 of the proposed model
90 with other existing models.
297

Table 5. Classification
Table6. accuracy
Overall performance of attained with various feature optimization algorithms in the OCI-
the models.
LSTM, DNN and DBN.
Metrics DNN (%) DBN (%) OCI-LSTM (%)
Feature Optimization Sensitivity 91.95 Accuracy of Accuracy of
96.03 Accuracy
98.78 of
Optimal Features (Sign)
Algorithm Specificity 90.54 DNN (%) 92.65DBN (%) OCI-LSTM
95.37 (%)
Correlation-based feature F18(f_3,
Score 90.73 94.77 87.32 97.32
f_7 to f_13) 81.82 88.45
selection (CFS)
Accuracy 91.72 95.73 97.11
Chi-squared (Chi_Sq) 11 (f_1 to f_4 and f_7 to f_13) 87.88 90.17 91.46
Genetic Algorithm (GA) Thef_4,
8 (f_3, comparative analysis
f_6 to f_10 and f_13) presented81.14
in Table 5 indicates
92.62a variation in94.51
the perfor-
Lease absolute shrinkage mance of the OCI-LSTM and the other two network models. The results obtained for all
and selection four
8 (f_2,metrics
f_3, f_5,of the
f_8, f_9,OCI-LSTM
f_11 to f_13)are superior when compared
84.51 to the DNN and94.7
91.46 DBN with
operator (LASSO) optimal features.
Ruzzo–Tompa (RT) From to
7 (f_3,f_7 Table
f_10,6,f_12
it isand
evident
f_13) that using90.57
the SSA, the classification
94.61 accuracy96.82
of the OCI-
LSTM has improved when compared with other famous feature optimization algorithms.
SSA 7(f_1, f_7, f_8, f_9, f_10, f_12 and f_13) 91.72 95.73 97.11
Figure 5 illustrates the comparison of the proposed model with other existing models.

Figure5.
Figure 5. Performance
Performance comparison
comparison of
of the
the proposed
proposedmodel
modelwith
withother
otherexisting
existingmodels.
models.

TableFigure
6. Classification accuracy
6 illustrates attained with of
the comparison various feature using
the results optimization algorithms
various in the opti-
well-known OCI-
LSTM, DNN
mization and DBN.applied to the OCI-LSTM, DNN, and DBN. The OCI-LSTM model
algorithms
demonstrates superior performance through rigorous experimentation and comparison
Accuracy Accuracy Accuracy of
Feature
with otherOptimization
models and techniques.Optimal Features
of DNN of DBN OCI-LSTM
Algorithm (Sign)
(%) (%) (%)
Correlation-based fea-
8(f_3, f_7 to f_13) 81.82 87.32 88.45
ture selection (CFS)
Chi-squared (Chi_Sq) 11 (f_1 to f_4 and f_7 to f_13) 87.88 90.17 91.46
8 (f_3, f_4, f_6 to f_10 and
Genetic Algorithm (GA) 81.14 92.62 94.51
f_13)
 Figure 6 illustrates the comparison of the results using various well-known optimi-
zation algorithms applied to the OCI-LSTM, DNN, and DBN. The OCI-LSTM model
Diagnostics 2024, 14, 239 15 of 19
demonstrates superior performance through rigorous experimentation and comparison
with other models and techniques.

100

90

80

70

60
CFS Chi_Sq GA LASSO RT SSA

Accuracy of DNN (%) Accuracy of DBN (%) Accuracy of OCI‐LSTM (%)

Figure6.6.Results
Figure Results comparison
comparison of
of the
the OCI-LSTM,
OCI-LSTM,DNN
DNNand
andDBN
DBNwith
withvarious
variousfamous optimization
famous optimiza-
algorithms.
tion algorithms.

TheOCI-LSTM
The OCI-LSTMstarts startswith
withaafeature
featureoptimization
optimizationstep stepusing
usingthe theSSA.
SSA.ThisThisstrategic
strategic
featureselection
feature selectionprocess
processaims aimsto toenhance
enhancethe themodel
modelperformance
performanceby byidentifying
identifyingthe themost
most
relevantattributes.
relevant attributes. Out Out ofof thirteen
thirteen features,
features, seven
seven optimized
optimized features
features (f_3,
(f_3, f_7,
f_7, f_8,
f_8, f_9,
f_9,
f_10,f_12,
f_10, f_12,andandf_13)
f_13)are areselected
selectedbased
basedon ontheir
theirstrong
strongcorrelations
correlationswith withthethetarget
targetvariable,
variable,
providing crucial
providing crucial insights
insights for predicting cardiovascular
cardiovascular disease.
disease. The TheSSA,
SSA,employed
employed in
in feature
feature optimization,
optimization, is designed
is designed to prevent
to prevent network
network generalization
generalization issues,
issues, such such as
as over-
overfitting. Overfitting
fitting. Overfitting occurs
occurs whenwhen a model
a model learns
learns fromfrom noise
noise in in
thethe training
training data,
data, leading
leading to
to poor
poor generalization
generalization onon unseen
unseen data.
data. By selecting
By selecting the most
the most suitable
suitable features,
features, the OCI-
the OCI-LSTM
LSTM mitigates
mitigates the risktheofrisk of overfitting,
overfitting, ensuring ensuring
robustrobust performance
performance on both ontraining
both training and
and testing
testing
sets. sets.
The
Theefficacy
efficacyofof thethe
OCI-LSTM
OCI-LSTM is evaluated
is evaluated through threethree
through comprehensive
comprehensive experiments.
experi-
In the first
ments. In theexperiment, the OCI-LSTM
first experiment, the OCI-LSTM is applied to thetoselected
is applied optimal
the selected features.
optimal In
features.
the second
In the second experiment,
experiment, thethe
same
same optimal
optimal features
features areareused
usedininthe
theDNN
DNNand andDeep
DeepBelief
Belief
Network
Network(DBN) (DBN)models,
models, andandthethe
results
resultsarearerecorded
recordedfor comparison.
for comparison. The third experiment
The third experi-
involves feeding the OCI-LSTM, DBN, and DNN with the
ment involves feeding the OCI-LSTM, DBN, and DNN with the original dataset contain-original dataset containing all
features. Additionally, the OCI-LSTM is tested with other feature
ing all features. Additionally, the OCI-LSTM is tested with other feature selection meth- selection methods. These
experiments provide a thorough
ods. These experiments provide aanalysis,
thorough highlighting the OCI-LSTM’s
analysis, highlighting effectiveness
the OCI-LSTM’s in
effec-
CVD prediction.
tiveness in CVD The OCI-LSTM
prediction. The is further enhanced
OCI-LSTM is furtherbyenhanced
integrating by the Genetic the
integrating Algorithm
Genetic
(GA) for optimizing the time window size for the LSTM units. This
Algorithm (GA) for optimizing the time window size for the LSTM units. This integration integration ensures that
the model adapts to the temporal aspects of the data, striking a balance
ensures that the model adapts to the temporal aspects of the data, striking a balance be- between capturing
essential information
tween capturing and avoiding
essential information overfitting.
and avoiding The overfitting.
GA-based network
The GA-based configuration
network
process contributes to the model’s overall efficiency.
configuration process contributes to the model’s overall efficiency.
The
The OCI-LSTM,
OCI-LSTM, particularly
particularly whenwhen combined
combinedwith withthe theSalp
SalpSwarm
Swarm and and
GA, GA, con-
consist-
sistently outperforms the other classifiers and conventional techniques. Comparative
ently outperforms the other classifiers and conventional techniques. Comparative evalu-
evaluations against the DNN, DBN, and models utilizing all the features showcase the
ations against the DNN, DBN, and models utilizing all the features showcase the superior
superior predictive power of the OCI-LSTM. The careful selection of optimal features and
predictive power of the OCI-LSTM. The careful selection of optimal features and the in-
the incorporation of evolutionary-based algorithms contribute to its remarkable perfor-
corporation of evolutionary-based algorithms contribute to its remarkable performance.
mance. In summary, the OCI-LSTM model stands out by addressing overfitting, opti-
In summary, the OCI-LSTM model stands out by addressing overfitting, optimizing fea-
mizing feature selection, and integrating the GA for improved temporal modeling. The
ture selection, and integrating the GA for improved temporal modeling. The experimental
experimental results and thorough comparisons emphasize its effectiveness in predicting
results and thorough comparisons emphasize its effectiveness in predicting cardiovascu-
cardiovascular disease.
lar disease.
The accuracy of the OCI-LSTM is achieved through a combination of effective feature
subset selection, resolution of network configuration challenges, iterative model enhance-
ment through optimization algorithms, and rigorous comparative analysis with other
models. The model’s ability to perform well with limited data further underscores its
potential for practical implementation in real-world scenarios.
Diagnostics 2024, 14, 239 16 of 19

The OCI-LSTM consistently outperforms the DNN and DBN across all metrics, sen-
sitivity, specificity, F1 score, and accuracy, showcasing its superior performance in CVD
prediction. The higher sensitivity and specificity of the OCI-LSTM suggest that it is effective
in minimizing both false positives and false negatives, making it a robust model for CVD
prediction. The high F1 score indicates a good balance between precision and recall in the
OCI-LSTM, reinforcing its effectiveness in capturing true positive cases while avoiding
false positives and false negatives. Overall, based on these metrics, the OCI-LSTM appears
to be a promising model for CVD prediction with a high level of accuracy and reliability.
The OCI-LSTM’s superior performance in CVD prediction, driven by effective feature
selection and optimized network configuration, empowers medical practitioners with a
reliable tool. The model’s high sensitivity and specificity ensure accurate identification
concerning both positive and negative cases, minimizing errors in diagnosis. Its ability
to maintain a balanced F1 score underscores its precision and recall, aiding practitioners
in capturing true positive cases while avoiding false positives and negatives. The OCI-
LSTM’s efficiency enhances practical applicability, providing medical professionals with a
robust and accurate predictive tool for timely intervention and personalized patient care in
real-world scenarios.

5. Conclusions
The Optimally Configured and Improved Long Short-Term Memory (OCI-LSTM)
model, integrating feature selection through the Salp Swarm Algorithm (SSA) and network
configuration optimization via the Genetic Algorithm (GA), adeptly tackles network gen-
eralization challenges like overfitting and underfitting. Through comparative analyses
involving the DNN and DBN models, utilizing both the complete feature set and the opti-
mized subset, the OCI-LSTM consistently emerges as the superior performer in terms of
accuracy. Rigorous statistical examinations and evaluations underscore the significance of
the OCI-LSTM compared to its counterparts. Achieving an impressive accuracy of 97.11%,
the OCI-LSTM holds substantial promise in supporting medical professionals in making
informed decisions for cardiovascular disease prediction. Furthermore, it involves collabo-
ration with domain experts to validate the significance of the chosen subset of features in
the context of cardiovascular health. Fine-tuning hyperparameters, such as the learning
rates and dropout rates, may also contribute to enhancing the OCI-LSTM’s robustness and
generalization across diverse datasets, and a user-friendly interface and integration into
existing healthcare systems could facilitate seamless adoption by medical professionals,
promoting its practical utility in real-world clinical settings.

Author Contributions: Conceptualization, T.K.R.; data curation, S.B. and S.D.; formal analysis, T.K.R.
and V.S.; investigation, T.K.R.; methodology, T.K.R.; project administration, S.B. and V.S.; resources,
S.D.; software, S.B. and S.D.; supervision, S.B.; validation, S.D.; visualization, V.S.; writing—original
draft, T.K.R.; writing—review and editing, S.D. All authors have read and agreed to the published
version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Online UCI Machine Learning Repository.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Polat, K.; Şahan, S.; Güneş, S. Automatic detection of heart disease using an artificial immune recognition system (airs) with fuzzy
resource allocation mechanism and k-nn (nearest neighbour) based weighting preprocessing. Expert Syst. Appl. 2007, 32, 625–631.
[CrossRef]
2. Trevisan, C.; Sergi, G.; Maggi, S. Gender differences in brain-heart connection. In Brain and Heart Dynamics; Springer:
Berlin/Heidelberg, Germany, 2020; pp. 937–951.
Diagnostics 2024, 14, 239 17 of 19

3. Uyar, K.; İlhan, A. Diagnosis of heart disease using genetic algorithm based trained recurrent fuzzy neuralnetworks. Procedia
Comput. Sci. 2017, 120, 588–593. [CrossRef]
4. Davie, A.P.; Francis, C.M.; Love, M.P.; Caruana, L.; Starkey, I.R.; Sutherland, G.R.; McMurray, J.J. Value of the electrocardiogram
in identifying heart failure due to left ventricular systolic dysfunction. BMJ Br. Med. J. 1996, 312, 222. [CrossRef] [PubMed]
5. Alam, M.M.; Saha, S.; Saha, P.; Nur, F.N.; Moon, N.N.; Karim, A.; Azam, S. D-care: A non-invasive glucose measuring
technique for monitoring diabetes patients. In Proceedings of International Joint Conference on Computational Intelligence; Springer:
Berlin/Heidelberg, Germany, 2020; pp. 443–453.
6. Vijayashree, J.; Sultana, H.P. Heart disease classification using hybridized ruzzo-tompa memetic based deep trained neocognitron
neural network. Health Technol. 2020, 10, 207–216. [CrossRef]
7. Chen, Y.; Chen, Z.; Li, K.; Shi, T.; Chen, X.; Lei, J.; Wu, T.; Li, Y.; Liu, Q.; Shi, B.; et al. Research of Carbon Emission Prediction: An
Oscillatory Particle Swarm Optimization for Long Short-Term Memory. Processes 2023, 11, 3011. [CrossRef]
8. Andreotti, F.; Heldt, F.S.; Abu-Jamous, B.; Li, M.; Javer, A.; Carr, O.; Jovanovic, S.; Lipunova, N.; Irving, B.; Khan, R.T.; et al.
Prediction of the onset of cardiovascular diseases from electronic health records using multi-task gated recurrent units. arXiv
2020, arXiv:2007.08491.
9. Paul, A.K.; Shill, P.C.; Rabin, M.R.I.; Akhand, M.A.H. Genetic algorithm based fuzzy decision support system for the diagnosis of
heart disease. In Proceedings of the 2016 5th International Conference on Informatics, Electronics and Vision (ICIEV), Dhaka,
Bangladesh, 13–14 May 2016; pp. 145–150.
10. Ashraf, M.; Ahmad, S.M.; Ganai, N.A.; Shah, R.A.; Zaman, M.; Khan, S.A.; Shah, A.A. Prediction of cardiovascular disease
through cutting-edge deep learning technologies: An empirical study based on tensorflow, pytorch and keras. In International
Conference on Innovative Computing and Communications; Springer: Berlin/Heidelberg, Germany, 2021; pp. 239–255.
11. Wiharto, W.; Kusnanto, H.; Herianto, H. Hybrid system of tiered multivariate analysis and artificial neural network for coronary
heart disease diagnosis. Int. J. Electr. Comput. Eng. 2017, 7, 1023. [CrossRef]
12. Liu, X.; Wang, X.; Su, Q.; Zhang, M.; Zhu, Y.; Wang, Q.; Wang, Q. A hybrid classification system for heart disease diagnosis based
on the rfrs method. Comput. Math. Methods Med. 2017, 2017, 8272091. [CrossRef]
13. Jin, B.; Che, C.; Liu, Z.; Zhang, S.; Yin, X.; Wei, X. Predicting the risk of heart failurewith ehr sequential data modeling. IEEE
Access 2018, 6, 9256–9261. [CrossRef]
14. Islam, M.Z.; Islam, M.M.; Asraf, A. A combined deep CNN-LSTM network for the detection of novel coronavirus (COVID-19)
using X-ray images. Inform. Med. Unlocked 2020, 20, 100412. [CrossRef]
15. Saha, P.; Sadi, M.S.; Islam, M.M. EMCNet: Automated COVID-19 diagnosis from X-ray images using convolutional neural
network and ensemble of machine learning classifiers. Inform. Med. Unlocked 2021, 22, 100505. [CrossRef] [PubMed]
16. Islam, M.M.; Karray, F.; Alhajj, R.; Zeng, J. A review on deep learning techniques for the diagnosis of novel coronavirus
(COVID-19). IEEE Access 2021, 9, 30551–30572. [CrossRef] [PubMed]
17. Islam, M.M.; Islam, M.Z.; Asraf, A.; Al-Rakhami, M.S.; Ding, W.; Sodhro, A.H. Diagnosis of COVID-19 from X-rays using
combined CNN-RNN architecture with transfer learning. BenchCouncil Trans. Benchmarks Stand. Eval. 2022, 2, 100088. [CrossRef]
18. Asraf, A.; Islam, M.; Haque, M. Deep learning applications to combat novel coronavirus (COVID-19) pandemic. SN Comput. Sci.
2020, 1, 363. [CrossRef] [PubMed]
19. Muhammad, L.J.; Islam, M.; Usman, S.S.; Ayon, S.I. Predictive data mining models for novel coronavirus (COVID-19) infected
patients’ recovery. SN Comput. Sci. 2020, 1, 206. [CrossRef] [PubMed]
20. Rahman, M.M.; Islam, M.; Manik, M.; Hossen, M.; Al-Rakhami, M.S. Machine learning approaches for tackling novel coronavirus
(COVID-19) pandemic. Sn Comput. Sci. 2021, 2, 384. [CrossRef] [PubMed]
21. Ali, L.; Rahman, A.; Khan, A.; Zhou, M.; Javeed, A.; Khan, J.A. An automated diagnostic system for heart disease prediction
based on chisquare statistical model and optimally configured deep neural network. IEEE Access 2019, 7, 34938–34945. [CrossRef]
22. Pourtaheri, Z.K.; Zahiri, S.H. Ensemble classifiers with improved overfitting. In Proceedings of the 2016 1st Conference on Swarm
Intelligence and Evolutionary Computation (CSIEC), Bam, Iran, 9–11 March 2016; pp. 93–97.
23. Latha, C.B.C.; Jeeva, S.C. Improving the accuracy of prediction of heart disease risk based on ensemble classification techniques.
Inform. Med. Unlocked 2019, 16, 100203. [CrossRef]
24. Tao, R.; Zhang, S.; Huang, X.; Tao, M.; Ma, J.; Ma, S.; Zhang, C.; Zhang, T.; Tang, F.; Lu, J.; et al. Magnetocardiography-based
ischemic heart disease detection and localization using machine learning methods. IEEE Trans. Biomed. Eng. 2018, 66, 1658–1667.
[CrossRef]
25. Arabasadi, Z.; Alizadehsani, R.; Roshanzamir, M.; Moosaei, H.; Yarifard, A.A. Computer aided decision making for heart disease
detection using hybrid neural network-genetic algorithm. Comput. Programs Biomed. 2017, 141, 19–26. [CrossRef]
26. Pérez, J.; Pérez, A.; Casillas, A.; Gojenola, K. Cardiology record multi-label classification usinglatent dirichlet allocation. Comput.
Methods Programs Biomed. 2018, 164, 111–119. [CrossRef] [PubMed]
27. Chatzakis, I.; Vassilakis, K.; Lionis, C.; Germanakis, I. Electronic health record with computerized decision support tools for
the purposes of a pediatric cardiovascular heart disease screening program in crete. Comput. Methods Programs Biomed. 2018,
159, 159–166. [CrossRef]
28. Mohan, S.; Thirumalai, C.; Srivastava, G. Effective heart disease prediction using hybrid machine learning techniques. IEEE
Access 2019, 7, 81542–81554. [CrossRef]
Diagnostics 2024, 14, 239 18 of 19

29. Ali, S.A.; Raza, B.; Malik, A.K.; Shahid, A.R.; Faheem, M.; Alquhayz, H.; Kumar, Y.J. 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
2020, 8, 65947–65958. [CrossRef]
30. Wang, L.; Zhou, W.; Chang, Q.; Chen, J.; Zhou, X. Deep ensemble detection of congestiveheart failure using short-term rr intervals.
IEEE Access 2019, 7, 69559–69574. [CrossRef]
31. Mirjalili, S.; Gandomi, A.H.; Mirjalili, S.Z.; Saremi, S.; Faris, H.; Mirjalili, S.M. Salp swarm algorithm: A bio-inspired optimizer for
engineering design problems. Adv. Eng. Softw. 2017, 114, 163–191. [CrossRef]
32. Hsiao, H.C.; Chen, S.H.; Tsai, J.J. Deep learning for risk analysis of specific cardiovascular diseases using environmental data and
outpatient records. In Proceedings of the 2016 IEEE 16th International Conference on Bioinformatics and Bioengineering (BIBE),
Taichung, Taiwan, 31 October–2 November 2016; pp. 369–372.
33. Abdeldjouad, F.Z.; Brahami, M.; Matta, N. A hybrid approach for heart disease diagnosis and prediction using machine learning
techniques. In International Conference on Smart Homes and Health Telematics; Springer: Berlin/Heidelberg, Germany, 2020;
pp. 299–306.
34. Gers, F.A.; Schmidhuber, J. Recurrent nets that time and count. In Proceedings of the IEEE-INNS-ENNS International Joint
Conference on Neural Networks, Como, Italy, 27 July 2000.
35. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrentneural networks on sequence modeling. arXiv
2014, arXiv:1412.3555.
36. Altan, G.; Kutlu, Y.; Allahverdi, N. A new approach to early diagnosis of congestive heart failure disease by using hilbert–huang
transform. Comput. Methods Programs Biomed. 2016, 137, 23–34. [CrossRef]
37. Hochreiter, S.; Schmidhuber, J. Long shortterm memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef]
38. Javeed, A.; Rizvi, S.S.; Zhou, S.; Riaz, R.; Khan, S.U.; Kwon, S.J. Heart risk failure prediction using a novel feature selection
method for feature refinement and neural network for classification. Mob. Inf. Syst. 2020, 2020, 8843115. [CrossRef]
39. Javeed, A.; Khan, S.U.; Ali, L.; Ali, S.; Imrana, Y.; Rahman, A. Machine learning-based automated diagnostic systems developed
for heart failure prediction using different types of data modalities: A systematic review and future directions. Comput. Math.
Methods Med. 2022, 2022, 9288452. [CrossRef] [PubMed]
40. Al Bataineh, A.; Manacek, S. MLP-PSO Hybrid Algorithm for Heart Disease Prediction. J. Pers. Med. 2022, 12, 1208. [CrossRef]
[PubMed]
41. Hassan, C.A.; Iqbal, J.; Irfan, R.; Hussain, S.; Algarni, A.D.; Bukhari, S.S.; Alturki, N.; Ullah, S.S. Effectively Predicting the Presence
of Coronary Heart Disease Using Machine Learning Classifiers. Sensors 2022, 22, 7227. [CrossRef] [PubMed]
42. Kurian, N.S.; Renji, K.S.; Sajithra, S.; Yuvasree, R.; Jenefer, F.A.; Swetha, G. Prediction of Risk in Cardiovascular Disease using
Machine Learning Algorithms. In Proceedings of the 2022 International Conference on Sustainable Computing and Data
Communication Systems (ICSCDS), Erode, India, 7–9 April 2022; pp. 162–167.
43. Rana, M.; Rehman, M.Z.; Jain, S. Comparative Study of Supervised Machine Learning Methods for Prediction of Heart Disease.
In Proceedings of the 2022 IEEE VLSI Device Circuit and System (VLSI DCS), Kolkata, India, 26–27 February 2022; pp. 295–299.
44. Islam, M.; Haque, M.; Iqbal, H.; Hasan, M.; Hasan, M.; Kabir, M.N. Breast cancer prediction: A comparative study using machine
learning techniques. SN Comput. Sci. 2020, 1, 290. [CrossRef]
45. Hasan, M.K.; Islam, M.M.; Hashem, M.M. Mathematical model development to detect breast cancer using multigene genetic
programming. In Proceedings of the 2016 5th International Conference on Informatics, Electronics and Vision (ICIEV), Dhaka,
Bangladesh, 13–14 May 2016; pp. 574–579.
46. Ayon, S.I.; Islam, M.M. Diabetes prediction: A deep learning approach. Int. J. Inf. Eng. Electron. Bus. 2019, 12, 21.
47. Haque, M.R.; Islam, M.M.; Iqbal, H.; Reza, M.S.; Hasan, M.K. Performance evaluation of random forests and artificial neural
networks for the classification of liver disorder. In Proceedings of the 2018 International Conference on Computer, Communication,
Chemical, Material and Electronic Engineering (IC4ME2), Rajshahi, Bangladesh, 8–9 February 2018; pp. 1–5.
48. Ayon, S.I.; Islam, M.M.; Hossain, M.R. Coronary artery heart disease prediction: A comparative study of computational
intelligence techniques. IETE J. Res. 2022, 68, 2488–2507. [CrossRef]
49. Smith, M.R.; Martinez, T. Improving classification accuracy by identifying and removing instances thatshould be misclassified.
In Proceedings of the 2011 International Joint Conference on Neural Networks, San Jose, CA, USA, 31 July–5 August 2011;
pp. 2690–2697.
50. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
51. Oh, M.S.; Jeong, M.H. Sex differences in cardiovasculardisease risk factors among korean adults. Orean J. Med. 2020, 95, 266–275.
52. Gers, F.A.; Schmidhuber, J.; Cummins, F. Learning to forget: Continual prediction with lstm. Neuralcomputation 2000, 12, 2451–2471.
[CrossRef]
53. Yadav, D.C.; Pal, S. Prediction of heart disease using feature selection and random forestensemble method. Int. J. Pharm. Res.
2020, 12, 56–66.
54. Srinivas, K.; Rani, B.K.; Govrdhan, A. Applications of data mining techniques in healthcare and prediction ofheart attacks. Int. J.
Comput. Sci. Eng. 2010, 2, 250–255.
55. Kim, Y.; Roh, J.H.; Kim, H.Y. Early forecasting of rice blast disease using long short-termmemory recurrent neural networks.
Sustainability 2018, 10, 34. [CrossRef]
56. Holland, J. Adaptation in Natural and Artificial Systems; University of Michigan Press: Ann Arbor, MI, USA, 1992.
Diagnostics 2024, 14, 239 19 of 19

57. Mitchell, T.M. Does machine learning really work? Aimagazine 1997, 18, 11.
58. Eiben, A.E.; Smith, J.E. Introduction to Evolutionary Computing; Springer: Berlin/Heidelberg, Germany, 2003; Volume 53.
59. Heitzinger, C. Simulation and Inverse Modeling of Semiconductor Manufacturing Processes; TU Wien Bibliothek: Wien, Austria, 2002.
60. Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv 2014, arXiv:1412.6980.
61. Haq, A.U.; Li, J.P.; Memon, M.H.; Nazir, S.; Sun, R. A hybrid intelligent systemframework for the prediction of heart disease
using machine learning algorithms. Mob. Inf. Syst. 2018, 2018, 3860146.
62. Available online: https://archive.ics.uci.edu/ml/datasets/heart+disease (accessed on 18 October 2021).
63. Maldonado, S.; Pérez, J.; Weber, R.; Labbé, M. Feature selection for support vector machines via mixed integer linear programming.
Informationsciences 2014, 279, 163–175. [CrossRef]
64. Alizadehsani, R.; Zangooei, M.H.; Hosseini, M.J.; Habibi, J.; Khosravi, A.; Roshanzamir, M.; Khozeimeh, F.; Sarrafzadegan, N.;
Nahavandi, S. Coronary artery disease detection using computational intelligence methods. Knowl.-Based Syst. 2016, 109, 187–197.
[CrossRef]
65. Yang, H.; Garibaldi, J.M. A hybrid model for automatic identification of risk factors for heart disease. J. Biomed. Inform. 2015,
58, S171–S182. [CrossRef]
66. Yahaya, L.; Oye, N.D.; Garba, E.J. A comprehensive review of heart disease prediction using machine learning. J. Crit. Rev. 2020,
7, 281–285.
67. Allen, L.A.; Stevenson, L.W.; Grady, K.L.; Goldstein, N.E.; Matlock, D.D.; Arnold, R.M.; Cook, N.R.; Felker, G.M.; Francis, G.S.;
Hauptman, P.J.; et al. Decision making in advanced heart failure: A scientific statement from the american heart association.
Circulation 2012, 125, 1928–1952. [CrossRef]
68. Vidhushavarshini, S.; Balasubramaniam, S.; Ravi, V.; Arunachalam, A. A hybrid optimization algorithm-based feature selection
for thyroid disease classifier with rough type-2 fuzzy support vector machine. Expert Syst. 2022, 39, e12811.
69. Valliappa, C.; Kalyanasundaram, R.; Balasubramaniam, S.; Sennan, S.; Kumar, N.S. Automated atrial fibrillation prediction using
a hybrid long short-term memory network with enhanced whale optimization algorithm onelectro cardiogram datasets. Int. J.
Non Commun. Dis. 2021, 6, 76.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.