Ischemic Heart Disease Prediction Using Optimized Squirrel Search Feature Selection Algorithm

In recent years, the volume in globally recognized medical data sets are increasing both with attributes and number of records. Machine learning algorithms aiming to detect and diagnose ischemic heart diseases requires high efficacy and judgment. The state of art Ischemic heart disease data sets presents several issues, including feature selection, sample size, sample imbalance, and lack of magnitude for some characteristics etc. The proposed study is primarily concerned with improving feature selection and reducing the number of features yet giving better decisions. In this study, to pick salient aspects of heart illness, an improved squirrel search optimization algorithm with a meta-heuristic approach is proposed. Comparison study of the proposed ischemic heart disease squirrel search optimization (IHDSSO) model in conjunction with random forest classifier ensures better feature selection and accuracy over 98% with respect to other state-of-the-art optimization algorithms.


I. INTRODUCTION
Today's generation is exceedingly full of activity in their daily routine agenda, which causes anxiousness, and stress. The heart rate and blood pressure of each individual are different, Pulse rates range from 60 to 100 beats per minute, while blood pressure ranges from 120/80 to 140/90. Worldwide, Heart Disease(HD) is the leading cause of death. ''Cardio'' means ''heart.'' Cardiologist disease is a category of heart disease. The various types of cardiac disease are Coronary artery disease, Myocardial infarction, Hypertrophic cardiomyopathy, Arrhythmia, Dilated cardiomyopathy, Heart failure, Mitral regurgitation, and Congenital heart disease.Nowadays, heart disease is the leading cause of death worldwide. Predicting heart disease is a complex task since it requires experience along with advanced knowledge [1], [2]. Various methods, such as the K-Nearest Neighbor Algorithm (KNN), Naive Bayes(NB), Decision Trees (DT), and Genetic Algorithm (GA), are used to categorize the severity of the The associate editor coordinating the review of this manuscript and approving it for publication was Rongbo Zhu . HD problem. Because the nature of the HD is complex, it must be treated with carefulness. Failure to do so could injure the heart or lead to death. Medical research and machine learning (ML) are being utilized to find various kinds of metabolic disorders. In the classification of HD and data analysis, ML with categorization plays a significant character [3].
''Data mining also referred to as knowledge discovery from data (KDD), is the automated extraction of patterns considering knowledge captured in large databases, the web, warehouses, other huge information repositories or data streams'' [4]. According to a world health organization (WHO) report [5], 17.9 million people died from cardiovascular disease (CVD) in 2016, accounting for 31% of all deaths globally. Stroke or heart attack accounts for 85 percent of these deaths. CVD fatalities account for three-quarters of all deaths, primarily in developing countries, low-and middle-income countries are affected. Of the 17 million deaths that occur too soon caused by non-communicable diseases in 2015, 82% were in countries with a low and moderate-income, and CVDs affected 37%. There is a need VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ to detect CVDs, and data mining algorithms can be used to achieve this [6], [7] A factor in data analytics that degrades productivity is missing data. An incorrect imputation may lead to incorrect prediction of missing values [8].
The best optimization algorithms in the world, known as meta-heuristics, have several benefits, including robustness, performance reliability, simplicity, ease of implementation, etc [9]. The goal of feature selection is to minimize the number of dimensions in our data, which may increase its generalization [10]. Because of the large amount of data, feature selection has become more important. Hybrid methods are more widely utilized in Feature Selection(FS) due to the computing difficulty of wrapper methods, filtering techniques' low performance, and embedding approaches' reliance on classifiers. Filtering metrics are used in hybrid techniques to minimize the computational complexity of wrapper algorithms, and they have been shown to provide better feature subsets. Although filtering metrics select features based on their importance, the majority of them are insecure and subjective toward the metric in question. Furthermore, the selection of filtering metrics is heavily influenced by the data distribution and data kinds [11] squirrel search algorithm (SSA) is an FS algorithm, a nature-inspired metaheuristic method, which was introduced in the year 2019. The goal of SSA is to emulate southern flying squirrels' dynamic foraging and gliding behavior. These behaviors are mathematically represented to emphasize exploitation and exploration in a specific search space [12]. An improved squirrel search algorithm (ISSA) was introduced to increase the efficiency and robustness of SSA. The progressive search approach and the leaping search method, both of which have an independent winter search strategy during the early evolutionary period, are included in the ISSA algorithm. The results of the experiments suggest that the ISSA algorithm can increase convergence accuracy, speed up convergence, and maintain population variety. In comparison to SSA, the statistical test shows that ISSA has substantial advantages [13]. The dataset information is obtained from Cleveland UCI heart disease and contains 303 instances with 76 attributes, out of only 13 of them are taken into account. Because they contain significant healthcare records, all of the features are important. The selection was based on the recommendation of a cardiac department expert. Clinical data are essential for determining the severity of cardiac disease and diagnosing it.
Recently, optimization has become even more crucial to our daily lives. Popular and widely utilized in many engineering domains are population-and evolutionary-based optimization techniques. This optimization approach intelligently selects the finest of the many possibilities accessible and offers a conducive setting for problem-solving. The majority of optimization techniques demand a mathematical model of the system [14]. It might be challenging to create a mathematical model for complicated systems. Even if the model is established, the high cost prevents using the solution time. To acquire adequate global and local search operators, it is difficult to build an optimization algorithm because of physical events [15].

II. RELATED WORKS
Jijesh et al. [16] created a supervised learning-based decision support system for multisensory healthcare data from wireless body sensor networks. The SSA is used to select features. Other methodologies such as Conventional Neural network, Deep Belief Network, and Artificial Neural Network are compared to the proposed M-DBN (modified deep belief network) technique. The proposed method outperforms all other techniques currently in use. Based on data mining techniques, Shan et al. [17] suggested a more accurate and practical risk prediction system. To reduce dimensionality, CFS Subset Evaluation was used with the Best-First-Search approach to picking key characteristics. The data is taken from the Cleveland Heart Disease Database and PKU People's Hospital's Cardiology inpatient dataset. SVM takes twice as long as the random forest classifier.
To pick salient aspects of heart disease, an imperialist competitive algorithm with a meta-heuristic approach is proposed in this paper [18]. In comparison to other optimization algorithms, this approach can deliver an additional best solution for FS toward genetics. The categorization is also done using the K-nearest neighbor approach. The efficiency of the FS method improved by applying the proposed algorithm, according to the evaluation results. The data for these tests came from the UCI ML repository's HD data and the Tehran Shahid Rajaei hospital. The method proposed by Fuad et al. [19] is feature optimization, which aims to find the better variable size n-gram attributes for the supervised ML, also known as ''discrete weights based n-gram feature selection.'' Preprocessing, appropriate attribute selection, attribute FS, and the classification approach used to predict the extent of the disease in a specific patient's medical records are all things to think about.
An Internet of Medical Things framework for the analysis of HD based on modified salp swarm optimization (MSSO) and an adaptive neuro-fuzzy inference system (ANFIS) is presented to increase prediction accuracy. The Levy flight method, the proposed MSSO-ANFIS improves search abilities. For all iterations, the suggested Levy-based crow search algorithm for FS acquired the maximum fitness values. When compared to existing approaches, the suggested MSSO-ANFIS strategy provides superior F1-score, precision, accuracy, and recall as well as the lowermost classification error [20].
Mohit et al. [12] introduce the SSA, a revolutionary natureinspired optimization paradigm. Glid-ing is a powerful mechanism that small mammals employ to traverse large distances. In terms of convergence rate and optimization accuracy, thorough comparative research is conducted to demonstrate the efficiency of SSA over other familiar optimizers. To test its applicability and resilience, the suggested technique is applied in real-time HD research. In comparison to other existing optimizers, the final results show that SSA gives better accurate solutions with a higher convergence rate.
For biomedical datasets, a robust filtering approach based on rank aggregation was developed in a hybrid FS model using the ISSA. Experiments on 9 biomedical data sets and three distinct classifiers demonstrate the resilience of our proposed model. The suggested technique beats previous models in terms of classification accuracy and computing time, according to the experiments. It has also been demonstrated that it reduces dimensionality to a greater extent [16]. Yanjiao et al. [13] proposed an ISSA. The suggested algorithm includes two searching methods: one is a jumping search, and the other is a progressive search. The SSA practical technique is selected automatically using a linear regression selection strategy, which improves the algorithm's robustness. The 'escape' operation in the jumping search method sufficiently develops the search space, while the 'death' operation further explores the developed space, balancing SSA's development and exploration abilities. The results of the experiments suggest that the proposed algorithm can increase convergence accuracy, speed up convergence, and maintain population variety.
An enhanced form of SSA called ISSA is suggested to improve global convergence capabilities. To begin, an adaptive predator presence probability technique suggested the algorithm's exploitation and exploration capabilities. Second, the randomness and fuzziness of flying squirrel foraging behavior are explained using a normal cloud model. The selection method between succeeding spots is implemented to uphold the greatest position of flying squirrel individuals. Finally, a dimensional search enhancement method is used to advance the algorithm's local search capability. When compared to advanced optimization methods like SSA, PSO, FOA, IFFO, and CMFOA, experimental test results show that ISSA performs well [21].
Kader et al. [22] have compared five metaheuristic algorithms used in the Covid19 team formation problem. Based on the ideal findings the crow search algorithm was the effective metaheuristic method. The modified Huber independent component analysis-based SSA [23] effectively reduce data dimensionality by combining modified Huber independent component analysis and the SSA. In an Infrastructure as a Service cloud environment, Sanaj et al. [24] propose a chaotic SSA(CSSA) for efficient multitask scheduling. The early ecosystem was developed with chaotic optimization for the efficient ecosystem to permit more global convergence. To enable the exploring authority to complement SSA methods, the suggested CSSA was finally produced the messy limited search.
Mesfer et al. [25] introduce a unique ''multi-objective squirrel search optimization algorithm with a stacked autoencoder'' (MOSSA-SAE)model for Financial Crisis Prediction in an IoT background. The MOSSA-SAE technique includes several sub-processes, including pretreatment, processing, and analysis. Handling class imbalances, parameter tuning, and categorization The MOSSA-SAE approach, for example, allows IoT devices such as smartphones to communicate with each other. Users' financial information is collected through computers, tablets, and other devices. For additional analysis, the data is sent to the cloud. Ponnuvel et al. [26] have proposed a study to handle the single area multi-fuel economic dispatch (SAMFED) and MAMFED problems, a Fuzzified SSA (FSSA) is proposed in this study. For unconstrained and limited portfolio optimization issues, Mahdi et al. [27] designed and adapted an SSA. For portfolio optimization, the suggested SSA metaheuristic technique is compared to a range of approaches reported in the works, including conventional single metaheuristics, hybrid metaheuristics, and multi-objective optimization approaches. The comparative analysis and computational findings utilizing various performance evaluations reveal that the suggested method is superior to both extended Mean-Variance and Sharpe models for unconstrained portfolio optimization. The suggested method has also produced very competitive results for the diverse methods used in the limited form of the issue. The RFE (recursive feature elimination) technique is used to select features. [28] [18]. Senthilkumar et al. [1] provide a unique ML-based model for selecting relevant features, which increases the accuracy of cardiovascular disease(CVD) classification. The prediction model is shown using a variety of feature combinations and many well-known classification strategies. Through the HD classification model using a hybrid RF and a linear model. Achieve a higher level of performance with an accuracy rate of 88:7%.
The Relief and Least Absolute Shrinkage and Selection Operator approaches are used to find suitable features [29]. To allow for comparisons, the findings are presented individually. The most accurate results were obtained by applying the relief and LASSO model. Jasmine et al. [30] did a study on evolutionary computing-based feature selection strategies, their advantages and disadvantages, and their contribution to CVD classification. According to the understanding study, having a larger number of features can lead to misclassification and overfitting. Anuradha et al. [31] offered a Feature Importance ranking of 2 gradient boosting algorithms, XGBoost and CatBoost.Subsets of features were generated using the relevance rating of each character as a threshold. On these subsets, the classifiers Majority voting ensemble, XGBoost, and CatBoost were used to model them, and the feature subset with the maximum accuracy was found. When calculating the value of a subset of attributes, the correlation-based FS method analyses the individual classify capacity of each attribute as well as the degree of redundancy between them [32].

A. DATA SET
The dataset is a collection of connected data, with a report for each instance based on the data it represents, and an attribute for each attribute in the dataset. This study uses data from ''Cleveland, Switzerland, Long Beach, and Hungary'', as well as data from the UCI repository [33] and Kaggle [34] for data investigation. There are 76 features in the data set, 14 of which are quite beneficial in identifying heart disease. In most cases, the predictive class attribute comes last. The attributes' data set specifications are shown in Table 1.

B. SIGNIFICANCE OF THE ATTRIBUTE
Every single property in the UCI HD dataset is thought to be important for HD prediction, and the importance of each one has been labeled as a hyperparameter. Table 2 shows the importance of the property, description, domain value, and hyperparameter.

IV. OVERVIEW OF FEATURE SELECTION TECHNIQUES
The process of reducing the number of input variables in a model is called feature selection. The total number of input variables is minimized to reduce the cost of modeling and, and to increase the model's performance. Statistical-based FS methods are used to statistics the relationship between the input variable and the target variable, then selecting the input variables with the strongest relationship. Although the statistical measures utilized are influenced by the type of data in both the input and output variables, these processes can be swift and successful. As a result, when performing filter-based analysis, an ML practitioner may find it difficult to select an appropriate statistical measure for a dataset. Figure 1 depicts an overview of FS approaches. In supervised techniques, the wrapper, filter, and intrinsic methods are used. Filter-based FS approaches score the correlation or dependence between input variables using statistical measures, which can then be filtered to find the most important features. The data type of the input variable and the output or response variable must be carefully considered while developing statistical measures for FS. 1) An optimization problem's structure could be such that it is impossible to define how to discover the precise answer. 2) From the perspective of the decision maker, the comprehensibility of metaheuristic algorithms can be much simpler. 3) Metaheuristic algorithms can be utilized for learning as well as the process of determining the precise solution. 4) Typically, definitions created using mathematical formulas overlook the trickiest aspects of real-world problems. Using flawed data to determine model parameters might result in far higher errors than using a metaheuristic technique, which only provides a partial solution.

A. SQUIRREL SEARCH ALGORITHM
The SSA is a dynamic foraging movement of southern FLS in deciduous forests of Asia and Europe by gliding, a successful strategy utilized by tiny mammals for long-distance travel [21]. During the summer, the squirrels move their places by gliding from one forest tree to the next, looking for foodstuff. They can simply locate acorn nuts to meet their daily energy requirements. Then they start looking for hickory nuts that have been kept for the winter. During the winter, they become less active and rely on the storage of hickory nuts to meet their energy needs. Flying squirrels become busier as the temperature warms. The SSA is built on the foundation of the above-mentioned process, which is repeated and continues throughout the squirrels' life span. It is a population-based technique in which each squirrel forages for food by moving around a multi-dimensional search space. The distance of the food from the squirrel individual is equivalent to the fitness value of the objective function, and the placements of squirrels are considered various design variables. Individual squirrels in SSA change their places and move to better ones. It assumes that there are n squirrels in a deciduous forest and that each tree has only one squirrel. The three types of trees are considered to be present in the forest. Normal, Acorn, and Hickory are the three types of trees. The forest is meant to feature N trees, one of which is a hickory tree, no acorn trees, and the rest are ordinary trees with no food. For squirrels, the hickory tree is the finest foraging location. The SSA is a foraging behavior simulation of a squirrel. Each squirrel uses four mechanisms to shift its position: population distribution, dynamic feeding behavior, seasonal adaptive intelligence, and random relocation of individuals at the end of the winter. Squirrel Search Algorithm incorporates a seasonal monitoring condition which facilitates the advantages of better and more efficient search space exploration. In addition, SSA aids to obtain the overall outcomes that balance exploration and exploitation challenges and eradicate over-fitting issues. However, SSA shares certain drawbacks with other intelligent evolutionary algorithms, such as poor convergence accuracy and slow convergence speed.

1) KEY STEPS IN SSA
Based on the nutrition foraging behavior of FLS, the optimization of SSA can be theoretically characterized by the following steps.
Step 1 (Initialize the Parameters): ''The total number of iterations Im, population size NP, the total number of decision variables Nd, predator presence probability Ppp, scaling factor Fs, gliding constant Cg, and upper and lower bounds for decision variables Ud and Ld. These options are selected at the start of the SSA process''.
Step 2: Random Initialization of Flying Squirrels SSA, like other population-based algorithms, begins with a random location of flying squirrels. In a forest, there are a specific number of flying squirrels (FLS), and their location can be set. The initial position of each FLS in the forest is determined using a uniform distribution. The positions of the FLS are randomly initialized in the search algorithm as follows: (1) where rand() returns a random number with a uniform distribution in the range [0, 1].
Step 3 (Fitness Evaluation): The fitness of each FLS is determined by entering the values of the decision variable into a user-defined fitness function and computing the corresponding values. The fitness value of each FLS's position indicates the type of food supply it is looking for, such as optimal, normal, or no food source, and hence its chances of survival. By entering the values of choice variables into a fitness function, The location of an individual flying squirrel's fitness value f = (f1f2,. . . , fNP) is computed as follows: i = 1,2,. . . NP Step 4: Sorting, Declaration, and Random Selection After recording the fitness values of each FLS position, the array is sorted in ascending order. A FLS with a poor fitness value lives in the hickory nut tree. The next 3 finest FLS are thought to be on acorn nut trees and will eventually migrate to hickory nut trees. The FLS that are left should be on normal trees. Some squirrels are thought to be moving to the hickory nut tree based on random selection, presuming they've met their daily calorie needs. The surviving squirrels will make their way to the acorn nut trees. The presence of predators always affects the FLS' foraging behavior. Sort the index concerning ascending order. The fitness value of the FLS' locations is then used to rate the quality of food sources in ascending order: Step 5: Generate new locations through Aerodynamic Gliding In each instance, it is supposed that in the lack of a predator, The FLS glides and searches the forest effectively for its preferred meal, but in the presence of a predator, it is forced to do a little random walk to find a nearby hiding site. Creating a new location by gliding is: where dg is random gliding distance and R1 is a function that returns a value from the uniform distribution on the interval [0, 1], Cg is a gliding constant, and R1 is a function that returns a value from the uniform distribution on the interval [0, 1]. The new position is estimated using aerodynamic gliding, either with velocity or random values. Move the new places while limiting their lower and higher bounds.
Step 6: Stopping Criterion The function tolerance criterion is a popular convergence criterion that specifies a small but permitted difference between the last two outputs. The maximum implementation time is occasionally used as a pausing condition. In this experiment, most iterations are used as a stopping condition.

V. METHODOLOGY
The first step of the Experiment workflow with the UCI HD data set has been loaded, and it is now ready for preprocessing. The 13 attributes ''Age, Sex, Cp, Trestbps, Chol, Fbs, Restecg, Thalach, Exang, Oldpeak, Slope, Ca, and Thal'' are selected from the preprocessed data set of HD. The RF algorithm was used to improve the classification accuracy. The confusion matrix is used to assess the model's performance. The experiment workflow is shown in Figure 2. This phase aims to select significant attributes, which will be useful in the classification of Ischemic Heart Diseases. The importance of the 13 attributes of the UCI HD datasets can't be overstated because they contain essential clinical details. These medical records are essential for determining the severity of HD and diagnosing it.   Figure 3 shows the comparison of the existing Feature selection technique with the proposed ISSA method. The obtained significant attributes are given to an existing random forest classification algorithm to classify Ischemic Heart Diseases. The output of the confusion matrix is validated in terms of accuracy, Sensitivity, and Specificity, which will provide a more accurate classification result for the classification of Ischemic Heart Diseases.
Table details a typical confusion matrix. The mathematical background of various parameters associated with the confusion matrix are: Accuracy: the proportion of instances that are correctly classified.  Kappa: Kappa is another single-value metric designed to help the algorithm assess performance among an array of classifiers. Kappa is designed to compare the performance of  Figure 4 shows the overall system model for Ischemic Heart Disease Squirrel Search Optimization (IHDSSO). The input of the proposed model is taken from the benchmark UCI HD dataset with missing values. Firstly, the missing values in the data sets are recognized and imputed by using the Ischemic Heart Disease Multiple Imputation Technique. Algorithm 2 illustrates a fuzzy-based multiple imputation algorithm for missing data [36].To identify the important features, the imputed data is provided to the Improved Squirrel Search optimization as an input. To classify Ischemic Heart Diseases, the collected relevant features are fed into an existing random forest machine learning classification algorithm. The performance is analyzed by using a confusion matrix.

A. RANDOM FOREST(RF) ALGORITHM
In this work, a variety of ML methods are compared with the proposed IHDSSO. For the proposed IHDSSO model evaluation, the random forest classification technique is used. To improve prediction capability, RF employs the notion of bagging to aggregate many DT. Individual learners are trained independently in bagging. It uses replacement to generate several samples of data at random from the unique dataset, and each DT is trained on distinct models of data. During tree creation, features are also chosen at random. A majority vote is used to integrate predictions provided by various trees. Figure 5 depicts the Random Forest's operation. By optimizing parameters like as the total number of estimators, size of the smallest node, and amount of  characteristics utilized to separate nodes, the RF can be modified for higher accuracy.

VI. RESULTS AND DISCUSSION
We have used the spider (python 3.8) compiler for the experimentation, the associated parameters are shown in table 4. Set the population's position at random values concerning the lower and upper bounds. The figure 6 shows the lower Algorithm 2 Ischemic Heart Disease Multiple Imputation Technique [36] Input: X, dataset with missing attributes. Output: The dataset with imputed values. Begin 2: foreach y ∈ X and ContainsMissing(y) do N←getNearestNeighbors(y,k) 4: foreach a ∈ A and IsMissing(a(y)) do J1→0, J2→0 6: foreach z ∈ N do M←(Lower approximation+Upper approximation)/2 8: J1←J1 + M× a(z) J2←J1 + M 10: end if J2>0 then 12: a(y)←(J1/J2) else 14: a(y)← (z)/n end if 16: end if end 18: end end  and upper bounds for a uniformly distributed random number in the range of 0 to 1. Based on the k neighbours and error rate, calculate the fitness value for the input feature data for N samples. Obtain the global fitness value among the estimated fitness value for the population is shown in the figure 7.
The convergence curves of the fitness values illustrated in Figure 8 reveal that the suggested IHDSSO technique has a superior convergence effect than other optimization models. Table 5 shows the comparisons of different algorithms. It demonstrates that the Bagging and SMO procedures have nearly identical amounts and both have an accuracy of  90 percent. The accuracy of Nave Bayes is the lowest. The sensitivity of the Naïve Bayes approach exceeds its accuracy; It tends to diagnose normal people. The neural network is 87 percent accurate. The ICA-KNN is 91 percent accurate. The proposed method is 98 percent more accurate than other methods. The results of comparing the proposed feature selection approach to different methods are shown in Figure 9. The Nearest neighbour (KNN) with Genetic Algorithm(GA) is used for effective classification. Using the Random Forest algorithm, the proposed method utilized stronger and smarter features for classification. As a result, classification accuracy has improved with fewer characteristics, resulting in lower test costs and greater test accuracy in identifying ischemic heart diseases [18].
Furthermore, except for IHDSSO, the OANN(DBMRI-TLBO-ANN) beat all of the analyzed approaches, with 95.41 percent accuracy, 95.74 percent F-score, and 90.76 percent kappa, respectively. The provided IHDSSO model, on the other hand, achieves 98.38 percent accuracy,  98.32 percent F-score, and 96.74 percent kappa, respectively. Table 6 depicts the classification results of several models in terms of specificity, sensitivity, Accuracy, F-Score and Kappa. Figure 10 shows the categorization models' results in terms of sensitivity and specificity. The Random Tree(RT) model is proven to be poor in the classification process, with a sensitivity and specificity of just 74.41 percent and 71.11 percent, respectively. Furthermore, the J48 classifier model is found to perform somewhat better classification, with a 78.94 and 78.03 percent sensitivity and specificity, respectively. Furthermore, the NBTree model has attempted to outperform previous models by achieving sensitivity and specificity of 80. 24 Figure 11 shows the categorization models' results in terms of kappa, accuracy, and F-score. The RT model is proven to be ineffectual in the prediction process, with limited 72.94 percent accuracy, 75.30 percent F-score, and 45.38 percent kappa value, respectively. Furthermore,    [37].
Finally, a detailed comparative analysis with newly presented models is conducted to further validate the IHDSSO model's improved classification. Table 7 shows the comparison findings. It is guaranteed that the presented IHDSSO model outperforms all other approaches tested, with a maximum accuracy of 98.38 percent. As a result, it can be used in real-time diagnosis as a classification model. To further validate the efficacy of the suggested model, the AUC of ROC charts were used to compare its performance to that of other models.    XGBoost, Random forest, and Extra tree model ROC charts, respectively. As indicated in the figures, the proposed model's ROC curve AUC is 0.9838, the RF model's AUC is 0.8804, the ET model's AUC is 0.8831, and the Bayesian optimization XGBoost model's AUC is 0.9180. AUC chart outcome reveals the proposed IHDSSO model outperforms Bayesian optimization XGBoost, RF, and ET models in terms of accurcy [38].

VII. CONCLUSION
In the present study, a new approach to the Ischemic Heart Disease Squirrel Search Optimization feature selection algorithm is verified considering the UCI heart disease dataset. The proposed IHDSSO model classifies the most essential features which could be used as a strong predictor for heart disease classification. In the study, we verified that the features like ''Cp'', ''restecg'', ''oldpeak'', ''Ca'', and ''thal'' are more essential for the prediction of heart disease. The proposed IHDSSO model in conjunction with a random forest classifier ensured accuracy of over 98.38% and optimal heart disease prediction. The proposed IHDSSO model will be incredibly helpful in supporting healthcare solutions for diagnosing Ischemic heart disease. However, there is a scope for improving the present study in terms of convergence accuracy and convergence speed.