An Improved Grey Wolf Optimization Algorithm for

Heart Disease Prediction

Sihan Niu a,*, Yifan Zhou b, Zhikai Li a, Shuyao Huang c, and Yujun
Zhou d
a Department of Computer Science, Johns Hopkins University, Baltimore MD, USA
b Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore
MD, USA
c
School of Computer Engineering and Science, Shanghai University, Shanghai, China
d
School of Information Sciences, University of Illinois Urbana-Champaign, Champaign IL,
USA
{sniu7, yzhou222, zli264}@jhu.edu, [email protected], [email protected]

Abstract. This paper presents a unique solution to challenges in medical image

processing by incorporating an adaptive curve grey wolf optimization
(ACGWO) algorithm into neural network backpropagation. Neural networks
show potential in medical data but suffer from issues like overfitting and lack of
interpretability due to imbalanced and scarce data. Traditional Gray Wolf
Optimization (GWO) also has its drawbacks, such as a lack of population
diversity and premature convergence. This paper addresses these problems by
introducing an adaptive algorithm, enhancing the standard GWO with a sigmoid
function. This algorithm was extensively compared to four leading algorithms
using six well-known test functions, outperforming them effectively. Moreover,
by utilizing the ACGWO, we increase the robustness and generalization of the
neural network, resulting in more interpretable predictions. Applied to the
publicly accessible Cleveland Heart Disease dataset, our technique surpasses
ten other methods, achieving 86.8% accuracy, indicating its potential for
efficient heart disease prediction in the clinical setting.

Keywords: Neural Network, Heart Disease, S-Shape Function, Grey Wolf

Optimization

1 Introduction and Related Work

The use of backpropagation neural networks has recently become a critical strategy
to combat heart disease, a leading global health challenge responsible for immense
distress and mortality[1]. Various conditions like coronary artery disease, heart attacks,
arrhythmias, and heart failure fall under the umbrella of heart disease, affecting
populations across the globe. While classical methods for identifying and forecasting
heart disease are grounded in manual estimations using determinants such as age,
gender, genetic history, blood pressure, cholesterol metrics, and daily habits[2,3], the
2

surge in accessible medical data has paved the way for heightened precision in heart
disease prognosis using machine learning (ML).
Machine learning models stand out with their innate ability to sift through expansive
medical datasets, reveal concealed correlations, and distinguish intricate associations
between multiple risk factors and instances of heart disease[4,5]. By identifying
individuals at a heightened risk earlier, one can initiate preventive actions that might
mitigate the severity of heart disease, safeguarding individuals, their families, and the
broader society. Moreover, the dynamic learning nature of ML models lets them
evolve with fresh data, refining their forecasting accuracy with time[6]. This ongoing
learning cycle bolsters their proficiency in discerning nuanced patterns and
multifaceted links that play a role in heart disease onset and evolution. Armed with
these insights, medical professionals can personalize therapeutic strategies, ensuring
optimized treatment routes and minimizing redundant tests, thus curbing overall
healthcare expenditures.
Several ML methodologies, like logistic regression, support vector machine, KNN,
decision tree, random forest, extra trees classifier, gradient boosting, Naive Bayes, and
XGBoost, have demonstrated immense promise in forecasting and averting
cardiovascular ailments. Timely detection and intervention of heart disease can
substantially alleviate its global toll on both individuals and medical infrastructure. The
emphasis on predicting heart disease facilitates timely interventions and fosters a
proactive preventive approach, enriching patient-centric decisions[7,8,9]. In this
landscape of predicting heart disease, numerous groundbreaking algorithms have been
put to the test to amplify precision. A prime example is the hybrid algorithm, MLP-
PSO, introduced by Ali Al Bataineh and Sarah Manacek[10]. Their trailblazing study
merges the strengths of MLP-PSO, offering an innovative way to fine-tune neural
network training.
In our study, we introduce an advanced neural network framework for precise heart
disease prediction, underpinned by an improved Grey Wolf Optimization algorithm.
Drawing from the predatory instincts of grey wolves, the ACGWO methodology fine-
tunes the neural network's coefficients and intercepts. Incorporating ACGWO within
the backpropagation mechanism, our composite model adeptly navigates the solution
domain, boosting the neural network's convergence rate and precision. The consistent
refinements in the ACGWO algorithm facilitate effective overall optimization,
detecting intricate data patterns associated with heart disease.

2 Method

2.1 GWO
The Grey Wolf Optimizer (GWO) is a bio-inspired optimization method, introduced
by Mirjalili et al.[11]. Drawing from the hierarchical social dynamics and predatory
patterns of grey wolves, the GWO aims to thoroughly navigate and pinpoint optimal
outcomes within a given problem domain. The procedure initiates by dispersing a
group of grey wolves randomly across the domain, with each wolf signifying a
possible solution. The methodology then progressively refines the wolves' positions
through three main phases: hunting, encircling, and preying. During the hunting phase,
 3

wolves modify their placements relative to the alpha, beta, and delta wolf
positions[12], maintaining a balance between exploration and exploitation to ensure a
holistic navigation of the potential solutions. As the optimization advances, wolf
positions undergo continuous adjustments, and their fitness is gauged against the
optimization problem's objective function. By methodically refining wolf positions and
leveraging the finest located solutions, the algorithm strives to identify the absolute
best outcome. While GWO has showcased its efficacy in multiple optimization
scenarios, it's pivotal to fine-tune parameters and make context-specific adjustments to
maximize its efficiency for distinct challenges.
2.2 Cauchy Distribution Using in GWO
The Cauchy distribution, named after the French mathematician Augustin-Louis
Cauchy, is a continuous probability distribution that often arises in various fields of
statistics and probability theory. It is also referred to as the Cauchy-Lorentz
distribution, as it is closely related to the Lorentzian function in physics. Based on the
previously mentioned grey wolf algorithm and Cauchy distribution, we have listed the
following Eq 1-4:
1 γ2
𝑓(𝑥; 𝑥0 , γ) = ⋅ 2 (1)
πγ γ + (𝑥 − 𝑥0 )2
𝑎2 1
fi_alpha = − ⋅ 2 ⋅ 𝑐2 + 𝑑2 (2)
π 2 Alpha_score
𝑎2 + ( − 𝑏2)
𝑓_𝑎𝑣𝑔
𝑎2 1
fi_delta = − ⋅ 2 ⋅ 𝑐2 + 𝑑2 (3)
π Delta_score
𝑎22 + ( − 𝑏2)
𝑓_𝑎𝑣𝑔
𝑎2 1
fi_beta = − ⋅ 2 ⋅ 𝑐2 + 𝑑2
(4)
π Beta_score
𝑎22 + ( − 𝑏2)
𝑓_𝑎𝑣𝑔
A distinguishing feature of the Cauchy distribution is its pronounced tails[13]. This
translates to a comparatively elevated likelihood of witnessing extreme outcomes than
in other distributions that have defined moments. As a result, the Cauchy distribution is
a go-to for representing data characterized by outliers or data exhibiting heavy-tailed
tendencies. It's essential to recognize that the Cauchy distribution lacks certain
fundamental properties, including the Central Limit Theorem. Owing to its absence of
finite moments, conventional statistical methods centered on mean and variance
estimations might be ill-suited for data adhering to the Cauchy paradigm. In spite of its
unique characteristics, the Cauchy distribution finds applications across multiple fields
such as physics, economics, and engineering. Whether utilized as a standard for
hypothesis evaluations or integrated within complex models, its significance is
palpable. However, while working with Cauchy-distributed data, one should proceed
with caution and consider methods specifically designed for handling heavy-tailed
distributions.
2.3 S-Shaped Curve
1
𝑓(𝑥) = (5)
1 + 𝑒 −𝑥
4

1
𝑤 = 𝑤start − (𝑤start − 𝑤end ) (6)
1 + 𝑒 (𝑎−𝑏⋅𝑡)
a 1
ww = ⋅ 2 ⋅c+d (7)
π iter
a2 +( − b)
Max_iteration

The S-Shaped curve, commonly known as the sigmoid curve, is frequently observed
in logistic regression models. It finds wide applications in various fields such as
population growth, plant growth, and virus spread.
The sigmoid and inverse sigmoid functions are shown in Eq 5 and Eq 6,
respectively. From Eq 5, it can be observed that as 𝑥 increases, 𝑦 also increases.
Initially, 𝑦 shows rapid growth, followed by slower growth in the later stages. In this
equation, an improved inertia weight linearly decreasing strategy for particle swarm
optimization is proposed.
The modified inertia weight formula is given as shown in Eq 6. The S-Shaped
curve, commonly known as the sigmoid curve, is frequently observed in logistic
regression models. It finds wide applications in various fields such as population
growth, plant growth, and virus spread. The Sigmoid function is expressed as where 𝑎
and 𝑏 are constants[14]. By setting 𝑥 = 𝑡, the expression for the inverse sigmoid
function is obtained. In Eq 7, 𝑡 denotes the ongoing iteration, while 𝑀𝑎x_i𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛
signifies the maximum iteration count for a particle. A reference study carried out a
comparative evaluation of four inertia weight functions: linear decrease, downward
parabolic curve, upward parabolic curve, and decreasing exponential curve. The
outcomes suggest that using the upward parabolic or decreasing exponential curve to
model inertia weight can amplify the optimization capability of the algorithm in
distinct measures.
Drawing from these findings, multiple strategies have emerged in recent times
aimed at bolstering the efficacy of particle swarm optimization. Examples include the
Gaussian function decreasing inertia weight approach, the logarithmic decrease
method, and the Cauchy distribution-based inertia weight decline technique[15-17].
While such modifications to inertia weight can indeed elevate the particle swarm
optimization's performance, they occasionally prompt premature convergence. This
happens due to an overly rapid decrease in inertia weight in the primary search phases,
leading to a diminished diversity amongst the particle population.
In a bid to counteract these challenges, our investigation taps into the attributes of
the inverse S-shaped function for inertia weight adjustments. As shown in Eq 7, this
function gradually diminishes across its span but exhibits slower shifts during the
commencement and conclusion phases. Such a modification ensures that the inertia
weight retains a relatively robust value in the initial phases, fortifying the algorithm's
global search potential. Conversely, in the latter stages, by sustaining a reduced inertia
weight for an extended duration, the algorithm’s local search prowess is augmented.
2.4 Training BP Using GWO
Merging the GWO algorithm with the training regimen of neural networks yields
several benefits. Primarily, the efficiency of GWO in navigating the search space
enables the neural network to bypass local optima, leading to superior solutions.
Moreover, the GWO's adeptness at striking a balance between exploration and
 5

exploitation aids the neural network in adapting well to novel data, thus minimizing
overfitting. Furthermore, with the GWO's inclusion, the training mechanism enhances
its resilience, lessening the odds of stagnation in less-than-ideal solutions. To sum up,
integrating Grey Wolf Optimization with neural network training rooted in
backpropagation presents a formidable method to bolster the network's efficacy in
navigating intricate optimization and pattern discernment challenges. By capitalizing
on the optimization prowess of the GWO algorithm, neural networks can witness
heightened precision and adaptability. This amalgamated strategy paves the way for
addressing intricate challenges across computer vision, pattern identification, and
beyond. Future studies can delve into refining and diversifying this integration,
propelling the discipline forward and addressing intricate real-world dilemmas.
2.5 ACGWO Algorithm for Neural Network Backpropagation
As shown in Table 1, the ACGWO algorithm addresses these issues by introducing
an adaptive strategy during the update of grey wolf positions. The Adaptive Curve
Grey Wolf Optimizer algorithm employs search agents (wolves) that iterate over the
search space to find the optimal solution. With each iteration, the positions of the
wolves are adaptively updated using calculated parameters such as 𝐴, 𝐶, and 𝐷, which
dictate their movement. The algorithm concludes by returning the position of the Alpha
wolf, representing the best-found solution.

Table 1. Pseudocode for the Adaptive Grey Wolf Optimizer

Algorithm Adaptive Curve Grey Wolf Optimizer

Require: Size of search space, number of search agents, maximum iterations
1: Initialize the position array Positions of the search agents (wolves)
randomly within the problem space.
2: for 𝑖𝑡𝑒𝑟 = 1 to 𝑚𝑎𝑥𝑖𝑚𝑢𝑚 𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 do
3: Calculate the fitness of each search agent, update the position and
fitness of 𝐴𝑙𝑝ℎ𝑎, 𝐵𝑒𝑡𝑎, and 𝐷𝑒𝑙𝑡𝑎.
4: for each 𝑊𝑜𝑙𝑓 in all 𝑊𝑜𝑙𝑣𝑒𝑠 do
5: Compute new 𝐴, 𝐶, 𝐷, 𝑤𝑎, and 𝑤𝑤 values
6: 𝐴 = 2 × 𝑤𝑎 × 𝑟1 − 𝑤𝑎
7: 𝐶 = 2 × 𝑟2
8: 𝐷 = 𝐶 × 𝑊𝑜𝑙𝑓_𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 − 𝐶𝑢𝑟𝑟𝑒𝑛𝑡_𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛
2
9: 𝑤𝑎 = 2 − 𝑖𝑡𝑒𝑟 × 𝑀𝑎𝑥_𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛
𝑎𝑎 1
10: 𝑤𝑤 = × 2 ×𝑐+𝑑
𝜋 𝑎𝑎2 +(
𝑖𝑡𝑒𝑟
−𝑏)
𝑀𝑎𝑥_𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛

11: Calculate the new position of the search agent

12: 𝑋_𝑛𝑒𝑤 = 𝑤𝑤 × 𝑊𝑜𝑙𝑓_𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 − 𝐴 × 𝐷
13: Trim the new position to the border of the search space if it
exceeds the search space
6

14: Update the position of the search agent with the new position
15: Update the position of the search agent with the new position
16: end for
17: Update the fitness of Alpha with the current best solution
18: end for
19: Output: Return the position of Alpha as the best solution found.

3 Experimental Results

3.1 The Comparison and Analysis of 9 Algorithms

The goal of this experiment is to compare the fitness of the ACGWO algorithm with
other well-known algorithms such as WDO, GSA, SCA, and PSO. We also compared
it to the traditional GWO algorithm, the standalone Adaptive GWO algorithm, and the
S-Shaped function GWO algorithm. The 6 basic test functions are depicted in the
following Table 2:

Table 2. Six benchmark test functions.

Function No. Function Name Range Optimal Value

f1 Sphere [−100, 100] 0
f2 Schwefel P2.21 [−100, 100] 0
f3 Schwefel P2.22 [−10, 10] 0
f4 Rosenbrock [−10, 10] 0
f5 Quadric Noise [−1.28, 1.28] 0
f6 Schafers [−100, 100] 0
In the experiment, we set the same population size as n+40, using search dimensions
of 30, 100, and 500. Learning factors were set as c1=c2=2. All other settings are
consistent with the original literature, including wmax=0.9 and wmin=0.4. The results
are presented in the following Table 3:

Table 3. The CGWO and AGWO are enhanced versions of the GWO algorithm. The CGWO
incorporates an S-shaped function, introducing a curve to the algorithm to expedite the
convergence rate. The AGWO is designed based on insights from the Cauchy inequality,
enabling each wolf in the algorithm to adaptively converge.

dim=30 dim=100 dim=500

Function Algorithm
MEAN STD MEAN STD MEAN STD

PSO 7.49E+02 5.37E+01 2.87E+03 1.78E+02 1.57E+04 3.83E+02

f1 SCA 1.52E+01 1.39E+01 2.65E+02 1.46E+02 3.18E+03 1.39E+03

GSA 3.75E+01 2.34E+01 3.05E+02 5.20E+01 1.82E+03 1.03E+02

 7

WDO 2.59E-07 3.83E-07 1.17E-05 1.38E-05 1.36E-04 2.07E-04

GWO 7.84E-06 5.45E-06 6.67E-02 3.93E-02 4.45E+01 7.47E+00

CGWO(Ours) 1.40E-92 3.16E-92 5.57E-88 4.28E-88 1.89E-84 1.80E-84

AGWO(Ours) 4.94E-131 1.08E-130 3.82E-127 1.25E-126 1.44E-124 1.10E-124

ACGWO(Ours) 6.01E-219 0.00E+00 5.63E-214 0.00E+00 2.21E-209 0.00E+00

PSO 7.08E+12 1.54E+13 1.03E+55 4.54E+55 3.56E+278 Inf

SCA 3.10E+00 2.79E+00 3.26E+01 1.52E+01 1.47E+02 7.45E+01

GSA 4.65E+01 9.90E+00 1.94E+02 1.88E+01 1.77E+271 inf

WDO 2.81E-03 3.51E-03 1.60E-02 1.30E-02 1.60E-01 1.23E-01

f2
GWO 3.55E-03 1.72E-03 8.65E-01 1.59E-01 6.74E+01 6.18E+00

CGWO(Ours) 5.51E-47 3.59E-47 4.34E-44 3.24E-44 6.53E-42 1.73E-42

AGWO(Ours) 1.14E-65 6.64E-66 1.18E-63 3.91E-64 9.77E-62 4.82E-62

ACGWO(Ours) 2.68E-109 5.79E-109 5.64E-107 4.67E-107 1.43E-104 1.49E-104

PSO 1.04E+03 1.53E+02 8.83E+03 1.06E+03 1.70E+05 5.01E+04

SCA 3.11E+02 1.78E+02 4.30E+03 1.67E+03 1.09E+05 1.95E+04

GSA 2.51E+02 9.93E+01 1.44E+03 4.51E+02 2.80E+04 1.19E+04

WDO 1.47E-05 2.15E-05 2.20E-04 3.84E-04 6.76E-03 1.03E-02

f3
GWO 1.58E+00 1.80E+00 3.95E+02 1.57E+02 1.33E+04 1.68E+03

CGWO(Ours) 5.26E-83 2.32E-82 1.28E-78 4.87E-78 1.54E-75 4.80E-75

AGWO(Ours) 5.21E-130 6.37E-130 2.00E-127 2.75E-127 4.71E-125 6.18E-125

ACGWO(Ours) 5.83E-218 0.00E+00 3.35E-215 0.00E+00 5.05E-211 0.00E+00

PSO 2.22E+01 2.46E+00 5.38E+01 4.56E+00 7.74E+01 3.85E+00

SCA 6.47E+01 9.63E+00 9.53E+01 1.85E+00 9.94E+01 2.45E-01

GSA 3.04E+01 6.17E+00 4.07E+01 4.11E+00 6.96E+01 1.98E+00

WDO 2.91E-03 2.07E-03 6.45E-03 5.40E-03 6.28E-03 5.30E-03

f4
GWO 7.97E-01 5.80E-01 3.65E+01 9.16E+00 8.10E+01 4.37E+00

CGWO(Ours) 2.66E-44 2.26E-44 7.04E-40 1.49E-39 3.36E-33 4.89E-33

AGWO(Ours) 4.15E-64 2.76E-64 7.18E-63 3.87E-63 1.62E-61 1.39E-61

ACGWO(Ours) 1.19E-107 1.28E-107 8.32E-106 1.30E-105 1.72E-104 1.31E-104

PSO 4.70E+02 1.82E+01 1.69E+03 5.43E+01 8.88E+03 1.17E+02

SCA 1.06E+02 4.87E+01 4.41E+02 1.35E+02 1.61E+03 5.21E+02

GSA 2.60E+02 3.83E+01 9.46E+02 8.19E+01 4.76E+03 1.51E+02

F5
WDO 1.22E+02 2.94E+01 5.19E+02 2.69E+02 8.75E+02 1.80E+03

GWO 2.67E+01 1.03E+01 1.88E+02 8.05E+01 1.71E+03 2.47E+02

CGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

8

AGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

ACGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

PSO 1.57E+00 9.66E-02 1.26E+01 1.92E+00 6.23E+02 3.34E+01

SCA 1.29E+01 9.41E+00 3.39E+02 1.80E+02 3.07E+03 8.47E+02

GSA 3.52E+02 5.50E+01 1.85E+03 1.25E+02 1.22E+04 3.01E+02

WDO 4.96E-02 1.54E-01 3.51E-04 4.80E-04 1.01E-03 2.60E-03

F6
GWO 3.51E-02 4.37E-02 1.01E+00 7.51E-02 4.18E+01 1.16E+01

CGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

AGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

ACGWO(Ours) 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00

 9

Fig. 1. The loss fitness value curve. (a) to (f) represent the loss fitness values of f1 to f6,
respectively, as the number of iterations changes at a dimension of 500.

Fig 1 illustrates that for both unimodal functions (f1-f4) and multimodal functions
(f5-f6), the ACGWO algorithm's standard deviation is zero. This data points to the
ACGWO algorithm's stability being superior to that of the other four adjusted
algorithms. Concurrently, during the tests on multimodal functions from f4 to f6, our
enhanced algorithm remains robust, even as the other five algorithms find themselves
trapped in local optima. Another observation from Fig 1 is the relatively consistent
evolutionary trajectory of the algorithm despite increasing iteration counts.
In side-by-side tests involving just the AGWO algorithm and the exclusive S-
Shaped function GWO approach that we included, it became evident that solely relying
on the adaptive curve expedites the convergence rate. However, introducing the S-
Shaped curve garners even better outcomes, most noticeably with the f5 function test.
Keeping in mind the frequent application of the f5 Rastrigin function in neural network
weight optimization, we plan to merge the ACGWO algorithm with neural networks in
forthcoming experiments and choose pertinent datasets for exploration.
3.2 The Comparison and Analysis of 9 Algorithms
In this section, the ACGWO has been applied to diagnose heart disease. The
Cleveland heart disease dataset has been used to evaluate the performance of the
ACGWO and other 6 models. Before constructing the models, the dataset undergoes
thorough preprocessing and cleaning steps, including addressing missing values,
converting categorical variables, and applying feature scaling using the standardization
method. By employing these techniques, we aim to enhance the accuracy and
effectiveness of the diagnostic models for heart disease. Table 4 below shows the
optimal parameter settings for Eq 2-4 and Eq 7, determined after multiple experimental
attempts.
Table 4. Parameters
GWO S-Shape Cauchy
Parameter Swarm Size Iterations a b c d a b c d
Value 100 1000 1.0 0.0 2.0 1.7 1.0 0.0 2.0 2.1
In Table 4, swarm size represents the number of individuals in the search space;
Iterations determine the number of iterations the algorithm will run.
3.3 Data Analysis
Exploratory Data Analysis (EDA) is a critical process of conducting preliminary
investigations on data. EDA is a significant step in fine-tuning the given dataset in a
different form of analysis to understand the insights of the key characteristics of the
dataset’s various entities. A visual representation of the correlation matrix of the
Cleveland heart disease dataset is shown in Figure 2:
10

Fig. 2. Correlationtrix for heart.csv

Figure 2 displays the pairs with the strongest correlation. Given that pairs with high
correlation essentially depict the same variance in the dataset, analyzing these pairs can
aid in identifying the more significant attribute within each pair for model
development. Intriguingly, no individual feature in the dataset exhibits a strong
correlation with the desired outcome. In addition, there are varied correlations among
the features in relation to the target value; some show a negative correlation while
others demonstrate a positive correlation, as illustrated in Eq 8.
∑𝑛𝑖=1(𝑥𝑖 − 𝑥̅ )(𝑦𝑖 − 𝑦̅)
𝑟𝑥𝑦 = (8)
√∑𝑛𝑖=1(𝑥𝑖 − 𝑥̅ )2 ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̅)2
3.4 Data Preprocessing
Data preprocessing is a fundamental step in machine learning, ensuring raw data is
refined and structured, paving the way for effective model training. Essentially, this
technique in data mining converts raw data into a more digestible and interpretable
format. In this study, our data preprocessing steps encompass: Managing Null or
Missing Values; Processing Categorical Data; Segmenting Data into Training and
Testing Sets; and Scaling Features.
3.5 Performance Evaluation Metrics
In the evaluation of our proposed diagnostic models for heart disease, we employed
rigorous performance metrics to assess their effectiveness. We use the same dataset as
Ali Al Bataineh and Sarah Manacek [10], a stratified 70-30 train-test split was
 11

employed, where 70% of the data was used for training, and the remaining 30% was
held out for testing across all models. The performance metrics used to evaluate the
effectiveness of the diagnostic models are as follows:
𝑇𝑃 + 𝑇𝑁
Accuracy = (9)
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
𝑇𝑃
Precision = (10)
𝑇𝑃 + 𝐹𝑃
𝑇𝑃
Recall = (11)
𝑇𝑃 + 𝐹𝑁
2 × Precision × Recall
F1 Score = (12)
Precision + Recall
3.6 Result and Discussion
The objective of this research was to craft a diagnostic model for heart disease
employing machine learning methodologies. Multiple models, with the ACGWO being
a notable mention, underwent training, fine-tuning, and assessment processes. Findings
from experiments conducted on the Cleveland heart disease dataset can be found in
Table 5.
Table 5. Performance Comparison of ML Classifiers[10]
Algorithms ACC AUC PRE Recall F1
Gradient Boosting 0.714 0.712 0.725 0.659 0.690
Decision Tree 0.758 0.756 0.775 0.681 0.738
Extra Trees 0.769 0.766 0.810 0.681 0.740
XGB 0.769 0.766 0.810 0.681 0.740
KNN 0.780 0.777 0.815 0.704 0.756
Random Forest 0.791 0.787 0.857 0.681 0.759
MLP 0.802 0.799 0.861 0.704 0.775
Logistic Regression 0.813 0.808 0.909 0.681 0.779
SVM 0.813 0.809 0.885 0.704 0.784
GaussianNB 0.824 0.821 0.868 0.750 0.804
MLP-PSO 0.846 0.848 0.808 0.883 0.704
ACGWO-BP(Ours) 0.868 0.863 0.854 0.880 0.870
In the above graph. We have found that ACGWO has shown the best results. That
might be due to several factorials that lead to this result. We have discussed this in
further context.
12

4 Conclusion
We initiated this research by addressing two main challenges faced by the GWO
algorithm: its proneness to local optima and its occasional slow convergence for
certain problems. Our approach to improving the convergence rate incorporated a
Sigmoid function, which led to noteworthy enhancements. An adaptive strategy was
adopted to combat the local optima challenge. Drawing inspiration from the
foundational principles of GWO, which replicates the dynamics of wolf packs, we
integrated a Cauchy distribution, chosen for its relevant physical implications. The
refined algorithm, when tested across six functions, showed superior performance
compared to the traditional GWO, evidencing a combined benefit. Our enhanced
algorithm, termed ACGWO, expanded the capabilities of the original GWO, especially
in equipment and detection use cases.
In the context of this research, we introduced an ACGWO-optimized BP neural
network tailored for diagnosing cardiovascular diseases. We delved into various
potential machine-learning models to craft intelligent cardiovascular diagnostic
systems. Using the Cleveland dataset, which comprises 13 distinct features to
determine the likelihood of heart disease, we conducted exhaustive experiments. Our
findings were illuminating: the proposed ACGWO-BP model surpassed all other tested
algorithms, achieving an impressive accuracy of 86.8%. This underscores the potential
of the ACGWO framework in aiding medical professionals to make more precise
diagnoses and prescribe optimal treatment pathways.

References
1. Al Bataineh, A.; Jarrah, A. High Performance Implementation of Neural Networks
Learning Using Swarm Optimization Algorithms for EEG Classification Based on Brain
Wave Data. Int. J. Appl. Metaheuristic Comput. 2022, 13, 1–17. [CrossRef]
2. Tsao, C.W.; Aday, A.W.; Almarzooq, Z.I.; Alonso, A.; Beaton, A.Z.; Bittencourt, M.S.;
Boehme, A.K.; Buxton, A.E.; Carson, A.P.; Commodore-Mensah, Y.; et al. Heart Disease
and Stroke Statistics—2022 Update: A Report From the American Heart Association.
Circulation 2022, 145, e153–e639. [CrossRef] [PubMed]
3. Petersen, K.S.; Kris-Etherton, P.M. Diet Quality Assessment and the Relationship between
Diet Quality and Cardiovascular Disease Risk. Nutrients 2021, 13, 4305. [CrossRef]
4. Samieinasab, M.; Torabzadeh, S.A.; Behnam, A.; Aghsami, A.; Jolai, F. Meta-Health
Stack: A new approach for breast cancer prediction. Healthc. Anal. 2022, 2, 100010.
[CrossRef]
5. Hameed, B.Z.; Prerepa, G.; Patil, V.; Shekhar, P.; Zahid Raza, S.; Karimi, H.; Paul, R.;
Naik, N.; Modi, S.; Vigneswaran, G.; et al. Engineering and clinical use of artificial
intelligence (AI) with machine learning and data science advancements: Radiology leading
the way for future. Ther. Adv. Urol. 2021, 13, 17562872211044880. [CrossRef] [PubMed]
6. Oikonomou, E.K.; Williams, M.C.; Kotanidis, C.P.; Desai, M.Y.; Marwan, M.;
Antonopoulos, A.S.; Thomas, K.E.; Thomas, S.; Akoumianakis, I.; Fan, L.M.; et al. A
novel machine learning-derived radiotranscriptomic signature of perivascular fat improves
cardiac risk prediction using coronary CT angiography. Eur. Heart J. 2019, 40, 3529–3543.
[CrossRef] [PubMed]
 13

7. Kumar, Y.; Koul, A.; Singla, R.; Ijaz, M.F. Artificial intelligence in disease diagnosis: A
systematic literature review, synthesizing framework and future research agenda. J.
Ambient. Intell. Humaniz. Comput. 2022, 1–28. [CrossRef] [PubMed]
8. Al Bataineh, A. A comparative analysis of nonlinear machine learning algorithms for
breast cancer detection. Int. J. Mach. Learn. Comput. 2019, 9, 248–254. [CrossRef]
9. Doppala, B.P.; Bhattacharyya, D. A Novel Approach to Predict Cardiovascular Diseases
Using Machine Learning. In Machine Intelligence and Soft Computing; Springer:
Berlin/Heidelberg, Germany, 2021; pp. 71–80.
10. Ali Al Bataineh, & Sarah Manacek. (2022). MLP-PSO Hybrid Algorithm for Heart
Disease Prediction. Journal of Personalized Medicine, 12(8), 1208.
https://doi.org/10.3390/jpm12081208
11. Mirjalili, S., Mirjalili, S. M., & Lewis, A. (2014). Grey Wolf Optimizer. Advances in
Engineering Software, 69, 46-61. DOI: 10.1016/j.advengsoft.2013.12.007
12. Gupta, S., & Deep, K. (2019). A novel Random Walk Grey Wolf Optimizer. Swarm and
Evolutionary Computation, 44, 101-112. DOI: DOI: 10.1016/j.swevo.2018.06.013
13. Ghosh, J., Li, Y. B., & Mitra, R. (2018). On the use of Cauchy prior distributions for
Bayesian logistic regression. Bayesian Analysis, 13(2), 359-383.
14. GU, Y., LU, H., XIANG, L., & SHEN, W. (2022). Adaptive Simplified Chicken Swarm
Optimization Based on Inverted S-shaped Inertia Weight. Chinese Journal of Electronics,
31(2).
15. RATNAWEERA, A., HALGAMAGE, G., & WATSON, H. C. (2004). Self-organizing
hierarchical particles warm optimizer with time-varying acceleration coefficients [J]. IEEE
Transactions on Evolutionary Computation, 8(3), 240-255.
16. CLERC, M., & KENNEDY, J. (2002). The particle swarm-explosion, stability, and
convergence in a multidimensional complex space [J]. IEEE Transactions on Evolutionary
Computation, 6(1), 58-73.
17. Lillicrap, T. P., Santoro, A., ..., & Hinton, G. (2020). Backpropagation and the brain.
Nature Reviews Neuroscience, 21(6), 335-346.