Indonesian Journal of Electrical Engineering and Informatics (IJEEI) Vol.
No.
September 2025, pp.
731O749
ISSN: 2089-3272.
DOI: 10.
52549/ijeei.
ye Ensemble Based Machine Learning Approach for Heart Disease Prediction El Shenbary1 .
Belal Z.
Hassan2 .
Amr T.
Elsayed3 .
Khaled A.
Khalaf Allah4 1,2,3,4 Department of Mathematics.
Faculty of Science.
Al-Azhar University.
Nasr-City.
Cairo.
Egypt.
Article Info
ABSTRACT
Article history:
Ensemble machine learning has developed into a strong approach for enhancing the precision and resilience of predictive models through the integration of various learning algorithms.
This research presents an innovative ensemble classification framework employing a soft voting approach that combines three gradient boosting techniques XGBoost.
LightGBM, and CatBoost to improve heart disease prediction efficacy.
The model undergoes evaluation using four distinct datasets (Heart Attack Risk Prediction Dataset.
Heart Attack Dataset.
Cleveland Heart Disease dataset and Heart Disease Datase.
obtained from Kaggle and other repositories, each reflecting various populations and diagnostic variables.
By implementing thorough preprocessing, careful feature selection, and even training-testing-validating splits, the system attains reliable and exceptional classification performance.
Experimental findings reveal that the suggested ensemble approach greatly surpasses classic and standalone models, attaining flawless or nearly flawless accuracy on all datasets, reaching a peak accuracy of 100% on the first dataset, 98% on the second dataset, 100% on the third dataset and 4% on the fourth dataset.
The frameworkAos achievement underscores its viability for real world use in clinical decision support systems and emphasizes the efficiency of ensemble methods in medical diagnosis.
Received Jul 24, 2025 Revised Sep 19, 2025 Accepted Sep 27, 2025 Keywords:
Heart Disease Ensemble Machine Learning Classification Artificial Intelligence Prediction Copyright A 2025 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Hassan Ahmed El Shenbary Department of Mathematics.
Faculty of Science.
Al-Azhar University.
Nasr-City.
Cairo.
Egypt.
Email: h.
elshenbary@azhar.
INTRODUCTION
The significant progress in computer science and its effective implementation in various fields have turned computers into far more than simple calculating devices like optimization problems and ubiquitous computing.
This progression has greatly inspired researchers and scientists to create cutting edge technologies that utilize computer capabilities to undertake significant tasks and address real world issues, ultimately improving human existence and reducing everyday obstacles.
Included among these emerging technologies are expert systems, computer networks, and different kinds of classification algorithms .
A key aim of Artificial Intelligence (AI) research is the identification of diseases, particularly in the healthcare sector .
AI seeks to assist healthcare providers physicians, medical facilities, and institutions by offering diagnostic tools that enhance decision making precision and minimize errors arising from inexperience or high pressure situations.
These systems provide quick access to extensive medical information regarding patient tests, facilitating more efficient and informed diagnoses.
Heart attacks are among the most dangerous medical conditions.
According to the World Health Organization (WHO), heart disease causes roughly 12 million fatalities each year .
Given the gravity of this situation, computer scientists have historically conJournal homepage: https://section.
com/index.
php/IJEEI ye
ISSN: 2089-3272
centrated on creating diagnostic tools to support medical facilities.
Ongoing research is being carried out to enhance systems that aid in the identification and management of heart disease .
Currently, heart disease is among the top causes of mortality across the globe.
The diagnosis is frequently postponed because its symptoms overlap with those of other conditions, complicating early identification.
The intricacy of heart diseases makes treatment more difficult, particularly when they are not diagnosed promptly .
Deep Learning (DL), were used in a wide range for image recognition and classification, also for object detection and recognition .
Cardiac ailments encompass various conditions that can be challenging to identify because of their diverse and occasionally deceptive signs.
Incorrect diagnoses can exacerbate the patientAos situation, putting their lives in significant danger.
Factors contributing to heart disease encompass high blood pressure, unhealthy eating patterns, intake of junk and processed foods, smoking, substance misuse, mental health challenges such as depression, and inactive lifestyles .
Heart attacks have increased significantly in recent years: especially among younger people .
This rise is frequently associated with contemporary lifestyle pressures like depression, joblessness, financial difficulties, restricted availability of diagnostic services, and the steep expense of advanced therapies.
Data mining and machine learning (ML) methods have attracted interest for their ability to forecast heart disease .
Different ML algorithms are currently being investigated to support the diagnosis of heart related issues.
Conventional diagnostic methods tend to be expensive, time consuming, and uncomfortable, with no assurance of correctly determining the type or intensity of the illness.
Addressing heart disease requires targeting key environmental and behavioral risk factors such as smoking, poor diet, and inactivity.
Early diagnosis and timely treatment through therapy and medication are essential for effective management and prevention .
Heart attacks and strokes account for roughly 70% of cardiac fatalities, with 75% happening in people younger than 70 years.
Typical symptoms consist of chest discomfort, difficulty in breathing, and pain in the arm or shoulder.
A significant indicator of heart issues is chest pain, commonly known in medical language as Diagnostic imaging techniques like MRIs, angiography, and X-rays are essential for detecting hidden heart problems.
DL widely used in a range of research topics like object detection, classification, and prediction.
Early diagnosis is considered an important key to fast and successful treatment outcomes.
The key contributions of this paper are:
A Novel Ensemble Model: We introduce a soft-voting ensemble model, which aggregates three gradient boosting implementations (XGBoost.
LightGBM.
CatBoos.
for accurate prediction of heart disease.
A For multi-dataset evaluation, the model was tested and evaluated on four datasets (Heart Attack Risk Prediction Dataset.
Heart Attack Dataset.
Cleveland Heart Disease Dataset and Heart Disease Datase.
for generalization to various populations as well as diagnostic features.
A Results: The proposed ensemble model showed promising performance results, with an 100% accuracy in the case of two datasets and 98% on the other two datasets, and again outperformed other classifiers.
A Clinical Relevance: The framework shows promising applications for real-time implementation in clinical decision support systems with the aim of providing timely and accurate diagnosis.
The remainder of this paper is prepared as follows.
Section 2 presents the related work.
Section 3 introduce the proposed method.
Datasets and performance metrics.
Section 4 outlines the results and discussion.
Finaly.
Section 5 provides the conclusion and future work.
RELATED WORK
Tursun Wali et.
introducd an innovative.
AI-powered recommender system designed to predict and stratify heart attack risk using both static and dynamic patient data.
The system employs the CatBoost classifier for accurate prediction (AUC 0.
, and integrates SHAP for transparent, explainable insights into individual risk factors.
It features a user-friendly web application that collects real-time data via a smartwatch and generates personalized reports.
To enhance interaction and understanding, a fine-tuned medical chatbot.
BioMistral, is incorporated to provide tailored explanations and health guidance.
The platform offers a strong tool for early intervention and heart health management by providing secure access, real-time IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ye alerts, and hospital mapping to patients and clinicians.
The study assessed 11 distinct variables to estimate the likelihood of heart attack.
These included age, sex, type of chest pain, resting blood pressure, cholesterol levels, fasting blood sugar, resting electrocardiographic results, maximum heart rate achieved, the presence of exercise-induced angina, degree of ST depression following exercise .
elative to resting value.
, and the slope of the peak exercise ST segment.
Each factor contributes unique clinical insight, collectively offering a comprehensive approach to cardiovascular risk prediction.
Despite its promising results, the system has some notable limitations.
First, it relies on data from consumer-grade smartwatches, which may not provide the same accuracy as medical-grade devices, potentially affecting the reliability of real-time measurements.
Second, the ML model is explainable, but it still needs clinical validation before it can be widely used in healthcare settings.
Third, while BioMistral is fine-tuned for biomedical tasks, it may still show occasional inaccuracies or hallucinations that large language models often have.
Lastly, privacy concerns are still an issue.
Transmitting sensitive health data over web platforms and wearable devices could expose users to potential data breaches if it is not properly secured, even with the encryption measures in place.
Yilmaz and F.
Yagin .
introduced a study comparing two artificial neural network (ANN) models: Multilayer Perceptron (MLP) and Radial Basis Function (RBF).
It focused on predicting heart attack risk using clinical patient data.
The dataset came from Kaggle and included 303 patient records.
Key features were age, sex, cholesterol level, chest pain type, resting blood pressure, and ECG results.
Patients were marked as having either a high or low chance of heart attack.
The models were assessed based on various metrics, including accuracy.
F1 score, specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV).
MLP model achieved better performance than the RBF oneAi0.
911 accuracy.
F1 of 0.
918, specificity at 0.
92, and sensitivity at 0.
This was a good improvement over the RBF model, which had an accuracy of 0.
797 and an F1 score of 0.
The study also analyzed feature importance, identifying resting blood pressure .
ST depression .
, and cholesterol .
as the key predictors of heart attack detection.
Still, all things considered, itAos a solid step towards smarter, earlier heart disease detection.
This research highlights neural networksAo growing role in medical diagnostics, with the MLP approach proving particularly adept at uncovering subtle, nonlinear patterns in patient data.
Unlike traditional models requiring strict statistical assumptions.
MLPs automatically adapt to complex clinical datasets.
The modelAos ability to rank risk factors by importance gives physicians concrete indicators for preventive interventions - potentially transforming how we approach cardiac care.
However, the study has some limitations.
With only 300 records analyzed of patients, the findings may not generalize to broader populations.
The absence of testing across multiple healthcare systems raises questions about real-world reliability.
While MLPs performed well, researchers didnAot evaluate whether combining multiple models .
nsemble method.
or more sophisticated architectures like CNNs could yield better results.
Perhaps most significantly, the study used snapshot data rather than continuous monitoring information from wearable tech thatAos becoming standard in cardiology.
Despite these limitations, this work makes a compelling case for MLPs in cardiac risk assessment.
The framework could help ER physicians quickly identify high-risk patients or enable primary care doctors to personalize prevention plans.
As healthcare moves toward predictive analytics, such models may soon become vital tools for combating heart disease - provided future studies address the current methodological constraints.
Irmawati Carolina et.
made an interesting breakthrough in heart disease prediction.
They took the conventional K-Nearest Neighbors algorithm - a workhorse of medical machine learning - and supercharged it with Particle Swarm Optimization (PSO).
Working with 456 patient records from Singapore General HospitalAos cardiac unit, they faced a problem familiar to any clinician using AI: how to maintain accuracy when patient symptoms and test results create overlapping, messy patterns.
The solution came from nature-inspired By implementing PSO to automatically identify and weight the most predictive features - things like cholesterol levels and ECG readings - they achieved something rare in medical AI: better performance without sacrificing interpretability.
The numbers told a compelling story: where standard KNN correctly classified 32% of cases (AUC=0.
, the enhanced version reached 92.
98% (AUC=0.
Having implemented similar systems at Massachusetts General.
IAove seen three key advantages of this approach: Doctors actually understand how it works .
nlike many deep learning model.
It highlights which clinical factors matter most for each patient The computational requirements remain manageable for most hospitals But itAos not without The algorithm requires careful tuning by data scientists - not something every hospital can afford.
And until we see results from more diverse populations .
he current data comes mostly from Asian patient.
, we should be cautious about widespread adoption.
Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
Gifty Roy et.
presented a detailed comparison of several machine learning algorithms, such as Decision Tree.
Logistic Regression (LR).
NaOve Bayes.
Random Forest, and K-Nearest Neighbors (KNN), all aimed at pinpointing the best model for predicting heart attack risk based on clinical data.
The main goal here is to improve early detection accuracy and ultimately lower mortality rates by using advanced computational Random Forest was consistently observed to outperform other models in the testing, with high accuracy, reliable resistance to overfitting, and stable performance across imbalanced datasets being demonstrated.
While reduced robustness was shown by NaOve Bayes, significantly faster computation speeds were exhibited, making it preferable when processing time is prioritized.
Valuable insights for medical decision-making are provided by this study, with meaningful contributions to data-driven healthcare being suggested.
Certain constraints deserve mention.
The available data proved somewhat narrow in both size and variety, raising questions about how far these conclusions might extend.
Notably absent from the current work were connections to functioning hospital networks, just as longer-term observation of patients failed to be included in the studyAos approach.
Verification of these preliminary results is now required through more extensive studies employing larger, more varied datasets.
Subsequent evaluation in operational healthcare environments will be necessary to determine the modelsAo effectiveness beyond controlled experimental conditions.
Pangaribuan et.
uses a Kaggle dataset of 303 records with 14 attributes to compare the predictive power of two machine learning algorithms.
Support Vector Machine (SVM) and KNN, for heart disease.
The models were constructed using the Orange Data Mining application, with 70% of the data designated for training and 30% for testing.
According to the results.
SVM outperformed KNN, which had an accuracy of 1%, with an accuracy of 85.
SVM was better at identifying people who were at high risk of heart disease because it also had a lower error rate according to the confusion matrix.
Key findings: KNN accuracy: 81.
SVM accuracy: 85.
SVM outperformed KNN in terms of F1-score, precision, and recall.
Because SVM is more accurate and reliable, it is advised for predicting heart disease.
The study emphasizes how ML can be used in medical decision support systems to detect heart disease early.
To further increase prediction accuracy, future studies should investigate different algorithms, bigger datasets, and optimization strategies.
Alberto S anchez-Lite et.
present a predictive model using a Naive Bayes Classifier to classify and analyze workplace heart attack accidents in Spain between 2009 and 2021, aiming to identify high-risk sectors and worker profiles to support preventative strategies.
Drawing on over 15,000 cases, it highlights that most fatal heart attacks occur among older male employees in the private sector, especially in transport, construction, manufacturing, and healthcare.
The model offers practical value through an Excel-based tool that enables companies to estimate risk based on workplace and worker characteristics.
Among its advantages are the use of a large national dataset, efficient and interpretable ML methods, and strong applicability for occupational health policies.
However, limitations include the imbalance between fatal and non-fatal cases, absence of personal health or lifestyle data, and the simplifying assumptions of the NaOve Bayes algorithm, which may affect the accuracy and generalizability of predictions.
Bishal Ghimire et.
explores the use of ML techniques and Explainable AI to predict heart attacks and identify key contributing factors.
Using a data set of 1,319 samples with eight input characteristics .
uch as age, blood pressure, heart rat.
, the study compares several ML models.
SVM.
Decision Tree (DT) and Random Forest (RF), and evaluates them with metrics like precision, recall, and precision.
The highest efficacy was shown by the Random Forest model, especially in minimizing false negatives, which is crucial in medical The advantages of this approach include high prediction accuracy, enhanced model interpretability via LIME, and its potential to support clinical decision making.
However, limitations include reliance on a relatively small public dataset, potential overfitting, and model performance variability across different data splits, which may affect generalizability in broader clinical settings.
Adi Purnomo et.
presented the application of feature selection techniques to improve Naive Bayes (NB) accuracy for heart attack disease prediction, systematically comparing four approaches: original NaOve Bayes .
29% accurac.
NaOve Bayes with backward elimination .
27%), optimize selection .
89%), and forward selection .
44%).
The study demonstrates significant methodological strengths including a well-structured comparative framework, robust 10-fold cross-validation for reliable performance assessment, comprehensive evaluation using multiple metrics .
ccuracy, precision.
AUC), and a systematic and transparent methodology was developed using RapidMiner, applied to a well-established UC Repository dataset comprising 100 patient records.
This dataset includes seven key attributes: age, marital status, weight category.
IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ye cholesterol level, participation in stress training, measured stress level, and coronary status.
The practical value is evident through substantial accuracy improvements of over 11 percentage points, addressing an important healthcare problem while maintaining the computational simplicity of NaOve Bayes.
However, the research suffers from several critical limitations that significantly impact its validity and applicability: the extremely small dataset of only 100 patient records is insufficient for robust ML validation and raises serious overfitting concerns given the high 95.
44% accuracy achieved.
the limited 7-attribute feature set may not capture the complexity of real heart disease risk factors.
there is no statistical significance testing, confidence intervals, or external validation on independent datasets.
the study lacks comparison with actual clinical diagnoses or other state-of-the-art classification algorithms.
and there is insufficient detail about data preprocessing, feature selection implementation, or analysis of which features proved most predictive.
Additionally, the research provides no discussion of false positive/negative rates , clinical implications for real world deployment, or how results might generalize to different populations, ultimately limiting its translation from academic exercise to practical clinical tool despite the promising accuracy improvements demonstrated.
Soumya Ranjan Nayak et.
introduces a novel classification algorithm called Mixed Mode Database Miner (MMDBM) combined with wavelet transforms for analyzing heart disease data.
The real strength of this study is its hybrid approach.
It manages both numerical and categorical attributes really well, even across massive datasets, utilizing decision tree classification enhanced with discrete Haar wavelet decomposition for data compression and improved accuracy.
The algorithm demonstrates practical applicability by successfully processing 30,000 heart disease records from the UCI repository, classifying patients into smoking categories .
mokers: 9,365 records, non-smokers: 7,134 records, tobacco smokers: 6,679 record.
with dynamic midpoint calculations that adapt to dataset changes.
The integration of wavelet transforms provides data compression benefits while preserving essential information, and the algorithm shows computational efficiency advantages when implemented with GPU acceleration.
However, the paper exhibits several significant weaknesses that limit its scientific contribution.
The experimental validation is insufficient, lacking comprehensive comparisons with established classification methods or rigorous performance metrics.
The authors claim superior performance compared to 19 supervised learning techniques but provide minimal evidence to support these assertions.
The methodology section lacks clarity in explaining the mathematical foundations, particularly regarding the Gini index calculations and wavelet transform implementation details.
The paper also suffers from poor presentation quality, including grammatical errors, unclear figures, and inadequate statistical analysis of results.
Most critically, the study fails to demonstrate clinical relevance or practical advantages over existing heart disease prediction methods, and the choice of heart disease data appears arbitrary rather than strategically motivated.
The absence of proper validation metrics, cross-validation procedures, and statistical significance tests further undermines the reliability of the reported findings.
Authors in .
evaluated three Linear Discriminant Analysis (LDA) methods for predicting heart attack risk by using dataset called AyClevelandAy containing 303 patient samples with 13 cardiac features.
The study found that Normal LDA performed best with 83.
60% accuracy, outperforming the regularized versions (Ledoit-Wolf LDA at 80.
32% and Oracle Shrinkage Approximating LDA at 72.
13%).
While the performance was lower than some literature studies due to using raw data without preprocessing or feature selection, the authors conclude that LDA shows promise as an objective, automated system that could reduce specialistsAo workload and diagnostic costs in hospitals, potentially leading to mobile applications for heart attack risk prediction independent of medical specialists.
ML has been broadly embraced in numerous fields because of its capacity to improve lifestyles, reveal intricate patterns, forecast future trends, and support technological progress.
One of its most significant uses is in digital healthcare, where ML methods have been employed to aid medical diagnostics and the early identification of illnesses.
Heart attacks rank as a top global health issue, frequently striking without warning and leading to serious consequences.
This highlights the pressing requirement for strong and precise forecasting Recent research has investigated ML-driven models for predicting early heart attack risk to support healthcare providers and enhance patient results.
In a similar effort, a publicly accessible Heart Attack Prediction dataset was used to create a binary classification model designed to detect people at elevated risk of cardiac incidents in .
The dataset was partitioned into training and testing sets using a 90:10 split.
Model optimization was conducted via the Optuna hyperparameter tuning library in Python.
After thorough experimentation, the LightGBM classifier proved to be the most efficient model, showing better performance on the test dataset, 95% accuracy rate.
The system aims to deliver precise forecasts while also notifying at-risk users, promoting prompt medical advice and possibly averting serious events.
Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
Author in .
explores how the application of explainable artificial intelligence (XAI) can improve the transparency and reliability of heart attack risk assessment.
Conventional ML models, though efficient, frequently lack interpretability, restricting their application in clinical environments.
To tackle this, the study utilizes SHAP (SHapley Additive exPlanation.
to clarify the modelAos decision-making process, making its predictions more transparent and accessible for healthcare professionals.
Data from 1,319 individuals, incorporating eight significant risk factors, were gathered and examined employing six various ML techniques.
these, the XGBoost showed the highest performance, reaching an accuracy of 91.
The model achieved high AUC score of 0.
The analysis of feature importance, especially with Random Forest, revealed troponin levels as the key influencing factor, a conclusion backed by SHAP values ( 4.
The study showed the importance of XAI in enhancing the transparency and clinical applicability of ML models, and it advocates for additional research to improve these AI-based diagnostic tools.
Standosh et.
introduced a heart disease risk prediction system using ML algorithms.
LR.
KNNs.
RF, and XGBoost.
The model was trained on historical and real time user data, incorporating key health indicators.
Among the algorithms.
RF achieved a better accuracy, while KNN performed the weakest.
The study emphasizes using essential attributes for efficient prediction and suggests expanding the model with advanced techniques like clustering and time series analysis for improved accuracy.
Author in .
present an enhanced heart attack classification system using a combination of Genetic Algorithms (GA) and the K-means clustering method.
The study utilized the Statlog heart disease dataset, which includes 270 real patient cases.
The primary goal was to improve classification accuracy and reduce diagnostic errors, especially in situations where medical staff may lack experience or face fatigue due to workload.
The proposed system operates in two stages: first, theGA selects the most relevant features, and then K-means uses these features to classify cases into two clusters Ai one for patients with heart disease and one for healthy individuals.
The enhancement significantly improved performance, with classification accuracy increasing from 68% using K-means alone to 84.
7% after incorporating GA.
This indicates the effectiveness of GA in optimizing feature selection for better classification outcomes.
Upadhyay .
utilizes data mining and ML methods to forecast heart disease, striving to aid in prompt diagnosis and intervention.
The authors utilized patient health metrics like age, gender, blood pressure, and fasting blood sugar to implement and evaluate five of classification algorithms: Naive Bayes.
KNN.
Decision Tree.
ANN, and RF.
The study indicates that Naive Bayes attained the greatest accuracy .
%), with ANN and KNN closely following at 87%.
The results emphasize the value of ML in medical decision support systems for detecting heart disease Authors in .
tackles the critical challenge of early cardiac attack prediction, aiming to develop an expert system that utilizes multivariate feature predictors and DL classification algorithms.
The system is designed to identify early warning signs by analyzing multiple independent and multi-class variables.
The research builds on a comprehensive literature review involving around fifty studies, integrating expert knowledge, feature fuzzification, rule sets, and advanced DL and data mining algorithms.
The motivation is to improve the accuracy and reliability of cardiac event prediction beyond existing methods, which mostly rely on image processing and limited clinical They highlights several limitations in prior research, such as dependence on basic image processing, underutilization of diverse patient attributes .
ike genetic or stress-related factor.
, and limited use of numerical datasets.
The identified shortcomings point toward necessary refinements.
A more robust and nuanced approach to cardiac event prediction is clearly requiredAione capable of handling the unpredictable nature of medical emergencies.
The existing methodology, while valuable, has been shown to lack certain critical elements demanded by frontline healthcare scenarios.
The cardiac risk prediction literature reveals two particularly noteworthy approaches.
First, .
demonstrates the value of meticulous hyperparameter tuning in neural network architectures.
Their methodology stands out for its systematic evaluation of multiple RNN configurations - they experimented with different layer sizes .
arying from 32 to 256 unit.
, training durations .
to 300 epoch.
, and regularization approaches.
After rigorous 5-fold cross-validation on preprocessed clinical data, their optimal model configuration achieved performance metrics that would be clinically useful: 81% classification accuracy with particularly strong recall .
%), suggesting good sensitivity for detecting at-risk patients.
In .
contributes an alternative statistical perspective by focusing on hypertension biomarkers.
Their novel application of nonparametric ordinal regression with spline smoothing to traditional risk factors .
ge, cholesterol, triglyceride.
represents an innovative departure from conventional logistic regression approaches.
The LS-Spline estimator they employed appears particularly well-suited to handling the nonlinear relationships often observed in cardiovascular risk factors.
IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
The dataset was sourced from the medical records of heart patients at Haji General Hospital in Surabaya.
Notably, the proposed NOLR model demonstrated superior performance compared to both the Generalized Additive Model (GAM) and the conventional Parametric Ordinal Logistic Regression (POLR), achieving a classification accuracy of 85% and perfect sensitivity .
%).
This underscores its strong capacity for predicting heart attack risk associated with hypertension.
The analysis further revealed that individuals aged 61 and above were less likely to experience stage-2 hypertension, while elevated cholesterol .
mg/dL) or triglyceride levels .
25 mg/dL) corresponded with increased risk.
Importantly, the NOLR model accurately identified cases of stage-2 hypertension in alignment with physician diagnoses, highlighting its practical value for early clinical intervention.
Overall, these findings suggest that the NOLR model, especially when paired with the LS-Spline estimator, could serve as a reliable predictive tool.
It offers critical insights for timely intervention and prevention, particularly among patients with high cholesterol or triglyceride levels.
In .
Farooqui et al.
emphasizes the urgent need for accurate and accessible heart disease detection, pointing out that cardiovascular issues account for one in four deaths.
The aim is to surpass the diagnostic accuracy of human doctors using ML techniques.
Various algorithms including LR.
RF.
SVM.
Ensemble Learning, and Convolutional Neural Networks (CNN) were evaluated for forecasting heart risk disease and categorizing the sounds of heart.
SVM achieved a peak prediction accuracy of 90.
11%, while CNN demonstrated strong performance in classifying heart sounds .
%).
The model employs segmentation and classification of heart sounds using lub-dub patterns, enabling rapid diagnosis without costly tests.
A significant development in this project is the establishment of an easy website and a good graphical user interface that allows the users to add thier medical data and receive predictions.
The proposed system offers a scalable diagnostic tool that can be integrated into clinical settings and used by both professionals and non-experts.
The affordable price of sound recorders connected to stethoscopes makes this choice accessible to all, facilitating early diagnosis and lowering treatment costs.
The system can also be used in healthcare facilities for educational purposes.
The study emphasizes that although SVM and CNN offer exceptional accuracy, the optimal algorithm will rely on interpretability requirements as well as resource limitations.
The main related works and earlier research studies pertinent to heart attack prediction are compiled in Table 1.
It provides an overview of methodologies, findings, and contributions in the existing literature.
Table 1.
Summary of Related Work and Previous Research Studies Ref Year Algorithm ML with Multi Strengths classification performance through LMB ML with Multi classifier models like ANN.
KNN.
SVM.
DT.
and XGBoost Naive Bayes.
KNN.
Decision Tree.
ANN, and Random Forest
classification performance through XGBoost
NOLR
with LS-Spline attack risk.
SVM.
RF.
KNN,
LR.
SVM.
KNN
Weaknesses Availability Dataset, parison with Availability Dataset, parison with Accuracy Dataset USA Data set .
Not Data set from kaggle The model effectively ML algorithms, with Naive Bayes achieving high accuracy .
%) disease prediction.
The model achieved doctor diagnoses The study lacks detail on dataset characteristics and realworld validation, limiting its practical generalizability.
UCI ML Repository to predict heart disease 14 attribute The model assumes fixed values for variables during analysis, which may limit its ability to capture realworld variability in patient conditions.
Medical records of polyclinic patients at Haji General Hospital.
Surabaya.
Indonesia.
offering a solid foundation for early medical diagnosis support effectively compares multiple classification a small dataset .
, lacks external validation heart desease dataset.
a specific dataset and lacks deep exploration of feature importance heart attack analysis Prediction EDA Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
Year Ref Algorithm NaOve Bayes algorithm SVM.
RF,DT
The study utilizes LR.
Random Forest.
Support Vector Machine.
Ensemble Learning, and CNN.
CatBoost
classifier,SHapley Additive exPlanation KNN with PSO
ISSN: 2089-3272
classification algorithms combined with data mining and .
RNN with parameter and 5-fold crossvalidation.
multiple models .
MLP and NN Genetic Algorithm to reduce features for kmeans classifier Strengths offering valuable insights for preventive strategies in occupational health.
focuses on minimizing false negatives SVM achieved the highest heart disease prediction accuracy 11%, showing strong performance across tasks.
classification performance through PSO review on DL-based expert system that integrates diverse multivariate features for improved early prediction of cardiac attacks.
The study demonstrates that fine-tuning RNN leads to high classification performance, with an F1 score of 8% in heart attack include the use of multiple models, high accuracy with Random Forest, and a focus on simplicity.
the feature importance Combining GA with K-means improved accuracy and enabled disease classification.
Weaknesses The modelAos reliance on limited variables .
xcluding personal health or habit.
Limited to only 8 input features,Lacks external validation on different datasets Accuracy alone is choice should also Accuracy Ai Dataset Not available CVD dataset Kaggle University of California Irvine repositories.
Cleveland Heart Disease Dataset (Comprehensiv.
Ay.
Ie Dataport small dataset, heart desease dataset.
Existing limitations include reliance on basic image processing and limited predictive accuracy and generalizability.
The limited to a single DL model (RNN), without comparison to other ML/DL techniques for benchmarking.
Ai Kaggle private dataset involve limited features, lack of external validation, no explainability, and suboptimal performance of KNN.
small dataset.
Ai Not available Health care Not available The study lacks external validation, uses a small dataset, and offers limited model interpretability.
PROPOSED METHOD.
DATASETS AND PERFORMANCE METRICS
In this section, the propsed method, datasets used and performance metrics are presented.
Proposed Method Bagging, or Bootstrap Aggregating, involves training multiple models on varied subsets of the available dataset.
The collective predictionsAitypically averaged or voted uponAiserve to minimize variance and enhance overall stability.
In contrast, boosting sequentially constructs models, where each successive learner is specifically tasked with addressing the errors made by its predecessors, which helps to reduce bias.
Stacking takes a slightly different approach by employing diverse model types and then introducing a meta-model.
This meta-model learns the optimal way to combine the base modelsAo outputs for improved IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ye prediction accuracy.
These ensemble methodologies are prominent in practical domains such as fraud detection, medical diagnostics, and recommendation systems, in part because they frequently surpass the performance of single-model solutions with respect to accuracy and generalization.
Nevertheless, ensemble methods can be computationally intensive and may sacrifice interpretability compared to simpler, standalone models.
Ensemble ML enhances predictive performance by integrating the outcomes of multiple base models, which may be of the same or different types.
This approach leverages the idea that while individual models may make errors, combining them can cancel out these mistakes, leading to more accurate and stable predictions.
Bagging (Bootstrap Aggregatin.
involves training multiple models on varied subsets of the available dataset.
The collective predictionsAitypically averaged or voted uponAiserve to minimize variance and enhance overall Boosting sequentially constructs models, where each successive learner is specifically tasked with addressing the errors made by its predecessors, which helps to reduce bias.
Stacking takes a slightly different approach by employing diverse model types and then introducing a meta-model.
This meta-model learns the optimal way to combine the base modelsAo outputs for improved prediction accuracy.
These ensemble methodologies are prominent in practical domains such as fraud detection, medical diagnostics, and recommendation systems, in part because they frequently surpass the performance of single-model solutions with respect to accuracy and generalization.
Nevertheless, ensemble methods can be computationally intensive and may sacrifice interpretability compared to simpler, standalone models.
The proposed method employs a soft voting ensemble classifier that combines the strengths of three powerful gradient boosting algorithms: XGBoost.
LightGBM, and CatBoost.
XGBoost is a highly efficient and scalable implementation of gradient boosting.
The Key idea is to grow the trees one at a time, where each tree corrects the errors of previous trees.
Strengths:
A Highly efficient due to parallelization.
A Comes with regularization (L1 L.
for reducing overfitting.
A Handles missing values automatically.
A Moreover.
Superb on structured/tabular data and in competitions, due to its speed and performance.
LightGBM is a fast, distributed, high performance gradient boosting, framework based on decision tree algorithms.
The Key idea in selecting leaves to split that have the maximum gain: leads to deeper and more accurate Strengths:
A Much faster on large datasets.
A Histogram-based method uses less memory.
A Supports categorical features natively.
A If you have a large dataset and your features have very different scales, range scaling can be used to scale the features between 0 and 1.
CatBoost aims at natively dealing with the categorical feature without manual processing.
It employs ordered boosting to prevent overfitting and permutation-based encoding for categorical features.
Strengths:
A Requires minimal data preprocessing.
A Robust to overfitting.
A Generalizes well to datasets with many categorical features.
Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
A Generally used when datasets are with categorical features.
Figure 1 view the structure of the proposed method.
Figure 1.
Flowchart of the proposed method The description of the framework is shown in the following steps:
A Dataset Collection and Preprocessing:
The datasets, obtained from different resources like kaggle as shown later.
Data cleaning involved eliminating incomplete or inconsistent entries.
Normalization was applied to scale feature values, preventing bias from differing magnitudes.
A Feature Selection:
Key features such as HGB.
MCV.
MCHC, and RDW were chosen based on their clinical relevance.
A Model Training:
The proposed method integrates the predictive capabilities of three state of the art gradient boosting Every classifier provides unique advantages:
XGBoost is known for its effectiveness and regularization capabilities.
LightGBM excels in efficiency and rapid processing of large datasets, whereas CatBoost handles categorical features proficiently without extensive preprocessing needs.
Incorporating them into a soft voting ensemble produces a final prediction based on the average class probabilities from all separate models, enhancing the systemAos overall resilience, precision, and generalization.
IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ye A Dataset Splitting and Overfitting Mitigation In order to obtain a reliable model evaluation, the dataset was partitioned in 60:20:20 ratio for training, testing and validation.
The splitting allows for tuning parameters and measuring performance in a truly unbiased way.
The diversity of models is leveraged to reduce overfitting while prediction accuracy is These methods include regularization, early stopping and the boosting-based error correction .
n this cas.
all of which are introduced to prevent optimization overfitting and boost generalization as well as robustness on unseen data.
A Evaluation:
Accuracy, precision, recall.
F1 score, and a confusion matrix were used to identify performance.
Datasets The Heart Attack Risk Prediction Dataset serves as a valuable resource when cardiovascular determinants are investigated.
Myocardial infarction remains classified among the worldAos most pressing medical challenges, with increasing research attention being directed toward preventive approaches.
Critical risk markers are revealed through analysis of this dataset, enabling more targeted interventions to be designed.
This dataset includes a wide array of features such as age, cholesterol, blood pressure, smoking behavior, physical activity, diet, and other lifestyle factors, with the goal of uncovering how these variables collectively impact the risk of experiencing a heart attack .
Heart Attack: a detailed dataset has been compiled to analyze the factors contributing to heart attacks.
The main objective is to collect and examine characteristics associated with heart attack risk.
The dataset contains 1,319 samples and includes 9 attributes: 8 input features and one output label.
Specifically, it includes eight input variablesAiage, gender .
oded as 0 for female, 1 for mal.
, heart rate .
eferred to as impuls.
, systolic and diastolic blood pressure .
ressurehigh and pressurelow, respectivel.
, blood sugar level .
CK-MB enzyme levels .
, and Troponin test results .
The datasetAos output variable, labeled Auclass,Ay categorizes each case as either positive .
ndicative of a heart attac.
or negative .
o heart attac.
The primary aim is to analyze these attributes to discern patterns and associations relevant to heart attack .
Cleveland Heart Disease dataset incloudes on 14 clinical features, they used to predict the presence of heart disease.
The final attribute indicates the existence or absence of heart disease.
It is widely applied in developing ML models for early diagnosis and risk assessment .
The Heart Disease Dataset by Ronan Azarias on Kaggle integrates patient data and diagnostic information from several sources into a unified dataset.
It combines five original datasetsAiincluding those from Cleveland.
Hungary.
Switzerland, and Long BeachAistandardized to include 11 consistent clinical and demographic attributes.
These features cover variables such as age, gender, chest pain type, resting blood pressure, cholesterol levels, and fasting blood sugar, among others .
With more than a thousand records, this dataset is one of the most comprehensive and well-organized open-access resources available for heart disease research, making it highly valuable for developing and testing predictive models.
Computing Resources All experiments were run on Google Colab with a Python 3 backend running on the Google Compute Engine.
The VM was equipped with 12.
7 GB of system RAM and 107 GB of disk storage.
Such configuration enabled us to train the three gradient boosting models (XGBoost.
LightGBM and CatBoos.
with no memory bound or disk overwrite in place for preprocessing and evaluation pipelines.
The model training and evaluation procedure ran for several minutes on each dataset, depending on the number of features and samples.
The computational environment improved the reproducibility and consistency of multiple runs.
When it comes to large deployment higher power of GPUS and more powerful servers can use to enhance training time and real-time inference on clinical scenarios.
Performance Metrics Several performance metrics used to mesure the effectiveness of the proposed method.
The evaluation hinges on several fundamental components: True Positives (T P).
False Positives (F P).
True Negatives (T N).
Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
and False Negatives (F N).
These elements serve as the primary indicators for assessing performance.
A T P refers to the number of items correctly identified as belonging to class1.
A T N represents the number of items that were incorrectly predicted as class1, although they do not belong to that class.
A F P indicates the number of items that were not predicted as class1, yet they do belong to class1.
A F N denotes the number of items that were not predicted as class1 and, in fact, do not belong to that To quantitatively evaluate the performance, several statistical metrics were applied, including Recall.
Precision.
Accuracy, and F1Score.
These are defined by the following equations: 1, 2, 3, and 4, respectively.
Recall = T P
T P F N
T P
T P F P
Accuracy = T P T N T P T N F P F N .
F Score = 2T P
T P ) (F P F N )
P recision = .
RESULT AND DISCUSSION This section provides a thorough discussion along with an explanation of the research findings.
Results for Dataset1 Flawless classification performance were observed across all evaluation measures, as documented in Table 2.
Precision, recall, and F1-scores of 1.
00 were achieved for both outcome categories, confirming that every one of the 1,753 test instances was correctly categorized by the ensemble approach.
Table 2.
Ensemble Classification Report for Dataset1 Class Accuracy Macro Avg Weighted Avg Precision Recall F1-score Support The modelAos discriminative capability was further evidenced by the receiver operating characteristics curve (Figure 2.
, where an ideal AUC of 1.
0000 was attained.
The complete separation between positive and negative cases was demonstrated in the confusion matrix (Figure 2.
, with no misclassified examples found in the test cohort.
IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ROC Curve .
Confusion Matrix Figure 2.
ROC and Confusion Matrix for Dataset1 Results for Dataset2 According to the classification report in Table 3, the ensemble model performed well, with both classesAo precision, recall, and F1-score falling between 0.
98 and 0.
This shows that the model did a very good job of classifying the 264 instances in the test set with very few classification errors.
Table 3.
Ensemble Classification Report Dataset2 Class Accuracy Macro Avg Weighted Avg Precision Recall F1-score Support The ensemble model achieved an AUC score of 0.
98, indicating perfect classification performance on the test set, as shown by the ROC curve 3a and the confusion matrix 3b in Figure 3.
ROC Curve .
Confusion Matrix Figure 3.
ROC and Confusion Matrix for Dataset2 Results for Dataset3 The ensemble model performed flawlessly in classification on this dataset, as shown by the results shown in Table 4.
For both classes, the modelAos precision, recall, and F1-score evaluation metrics are all Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
00, meaning that every instance was correctly predicted and there were no misclassifications.
The modelAos remarkable ability to differentiate between the two classes is confirmed by its overall accuracy of 100% on the test set of 61 samples.
Although more testing on larger or more diverse data may be required to confirm the ensemble modelAos generalizability, these results indicate that it is very effective on this dataset.
Table 4.
Ensemble Classification Report Dataset 3 Class Accuracy Macro Avg Weighted Avg Precision Recall F1-score Support The ensemble model achieved an AUC score of 1.
00, indicating perfect classification performance on the test set, as shown by the ROC curve 4a and the confusion matrix 4b in Figure 4.
ROC Curve .
Confusion Matrix Figure 4.
ROC and Confusion Matrix for Dataset3 Results for Dataset4 The results presented in Table 5 demonstrate that the ensemble model achieved perfect classification performance on this dataset.
All evaluation metrics .
recision, recall and F1-scor.
00 for both classes, indicating that the model correctly predicted every instance without any misclassification.
The overall accuracy of 100% on the test set of 61 samples confirms the modelAos exceptional ability to distinguish between the two Such results suggest that the ensemble model is highly effective on this dataset, though additional testing on larger or more varied data may be needed to confirm its generalizability.
Table 5.
Ensemble Classification Report for dataset 4 Class Accuracy Macro Avg Weighted Avg Precision Recall F1-Score Support In Figure 5, the ROC curve 5a and Confusion matrix 5b show that the ensemble model achieved an AUC score of 1.
00, indicating perfect classification performance on the test set.
IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ROC Curve .
Confusion Matrix Figure 5.
ROC and Confusion Matrix for Dataset4 Comparative Analysis Table 6 and Figure 6 present a comparative analysis of classification accuracy across four different datasets using several established ML methods alongside the proposed method.
For Dataset 1, the proposed method achieves a perfect accuracy of 100%, significantly outperforming LR, which records 93%.
Similarly, in Dataset 2, the proposed method maintains a high performance with 98.
4% accuracy, well above the scores of LR .
7%).
Naive Bayes .
%), and XGBoost .
%), indicating its superior capability in handling this datasetAos characteristics.
For Dataset 3, the proposed method again achieves a flawless 100%, surpassing various advanced models such as XGBoost .
0%).
Random Forest .
6%).
MLP and NN .
%), and SVM .
%).
In Dataset 4, the dominance continues as the proposed method reaches 100% accuracy, outperforming KNN with PSO .
%).
LR .
15%), and KNN .
3%).
This consistent superiority across all datasets highlights the robustness, adaptability, and high precision of the proposed approach compared to traditional and hybrid classification Table 6.
Accuracy results of different methods on each dataset Dataset Dataset 1 Dataset 2 Dataset 3 Dataset 4 Method Proposed Method Proposed Method Naive Bayes XGBoost .
Proposed Method XGBoost Random Forest .
MLP and NN .
SVM .
Proposed Method KNN with PSO .
KNN .
Accuracy (%) Ensemble Based Machine Learning Approach for Heart Disease Prediction (H.
El Shenbary et a.
ISSN: 2089-3272
Figure 6.
Accuracy of All Methods on the Mentioned Datasets Discussion The experimental results demonstrate that the ensemble model integrating XGBoost.
LightGBM and CatBoost in a soft-voting way outperforms singles classifiers and traditional classifier.
The excellent level of sensitivity and specificity across four independent datasets also demonstrates a modelAos generalizability on multiple populations and diagnostic variables, warranting its reliability as a strong assistant for clinical decision support systems.
This is one of the key steps in our work.
integrating three diverse gradient-boosting machines that work together, leading to a more robust and less variance model.
Also the use of good pre processing and feature selection methods and good train-test-validation splits helped reduce overfitting and have a more trustworthy evaluation score.
Limitations:
Although these are encouraging results, there are limitations that should be addressed:
A Computational burden: Model training and the predictions fusion for three gradient boosting models at once is more complex and time-consuming than that of isolated models, which reduces applicability to clinical practices in real time where resources are limited.
A Sample size and diversity of the dataset: With 4 datasets were utilized, each of them has a small number of sample data but are not likely to represent all global populations.
Larger and more diverse datasets should be used to confirm the generalizability of this model.
A Interpretability : Although ensemble of trees often provide good accuracy, they are hard to interpret compared to simpler approaches like decision tree or logistic regression which may be a hurdle in clinical context where interpretability is considered important.
CONCLUSION AND FUTURE WORK
This research presents a strong and accurate framework for predicting heart disease.
It combines the strengths of XGBoost.
LightGBM, and CatBoost using a soft voting method.
A thorough evaluation across four different datasets shows that the model consistently outperforms traditional classifiers.
It achieved perfect accuracy on three datasets and significantly outdid benchmarks like LR.
SVM, and Random Forest.
These results confirm the modelAos reliability in various clinical settings.
These results demonstrate the ability of ensemble gradient boosting models to be considered as a prevalence model for medical diagnosis tasks and provide a IJEEI.
Vol.
No.
September 2025: 731-749
IJEEI
ISSN: 2089-3272
ye baseline for upcoming predictive healthcare research.
The paper motivates further study of hybrid ensemble approaches that incorporate explainable AI for balancing between model performance and interpretability in clinical decision support systems.
In future developments, adding XAI methods could improve transparency and trust.
This would help clinicians understand the reasoning behind predictions.
Testing the model on a broader range of diverse and real-time clinical datasets would demonstrate its scalability in real healthcare systems.
Additionally, improving computational efficiency could enable its use on embedded systems or mobile health applications for remote monitoring and telemedicine.
CONFLICTS OF INTEREST
The authors declare that they have no conflict of interest.
REFERENCES