Hybrid Machine Learning Model for Early Prediction of Coronary

Artery Disease: Integrating LightGBM and Ensemble Techniques for

Enhanced Accuracy

Kavin Kumar D., Devendhiran S., Gomathy G. and Karnish N.

Department of Computer Science and Engineering, Nandha Engineering College, Erode, Tamil Nadu, India

Keywords: Coronary Artery Disease, LightGBM, Ensemble Learning, Feature Selection, Class Imbalance.

Abstract: Coronary Artery Disease (CAD) is a major issue confronting the global community today, which cannot be

foregone without some accurate predictive models for early diagnosis and intervention. In this study,

therefore, a Hybrid Machine Learning Model, HY OptGBM-Ensemble, is designed to combine LightGBM

with techniques for ensemble aiming at higher accuracy in predictions while dealing with class imbalance.

The model uses Optima-based hyperparameter tuning along with focal loss optimization and recursive feature

elimination to fine- tune feature selection and improve classification. Comparative evaluation against

LightGBM, XGBoost, and Logistic Regression shows that the proposed hybrid model has obtained AUC

scores 97.8% on the Framingham dataset. Class distinction is also improved along with predictive capability.

SHAP analysis further increases model interpretability.

1 INTRODUCTION

An important cause of death is coronary artery

disease (CAD), resulting from the accumulation of

atherosclerotic plaques in the coronary arteries, which

decreases blood flow. Significant outcomes include

angina, heart failure, and myocardial infarction,

making early detection essential. Machine learning

(ML) has contributed to CHD diagnosis through risk

prediction models, imputation algorithms, and hybrid

classification models. These advances enable CAD

risk prediction and early intervention. In this paper, a

hybrid ML model combining LightGBM with

ensemble methods, with hyperparameter tuning and

advanced loss functions, is presented to enhance

prediction accuracy and early detection. It combines

recursive feature elimination for the best feature

selection and SMOTE for handling class imbalance to

realize better recall in minority class prediction.

SHAP analysis is also used to realize better

interpretability as a measure of better understanding

of significant risk factors on CAD. Experimental

verification shows that the suggested model is more

accurate, precise, and retains a greater number of

cases than conventional models, thus supporting its

use in clinical decision-making and the early

detection of diseases. The model surpasses

conventional classifiers, guaranteeing precise CAD

diagnosis via ensemble learning and optimization.

SHAP analysis improves interpretability, rendering it

a significant resource for clinical applications. This

interpretability empowers healthcare professionals to

understand the underlying factors influencing

predictions, facilitating better- informed decision-

making in patient care. By prioritizing early

detection, this model aims to reduce the burden of

CAD on healthcare systems and improve patient

outcomes. Its ability to adapt increases its usefulness

in clinical environments, while sophisticated machine

learning methods provide healthcare providers with

valuable insights for improved management of CAD.

2 RELATED WORKS

Advanced machine learning techniques have been

used to predict early-stage Coronary Artery Disease.

Ensemble learning, boosting algorithms, and

optimization methods have been applied. Since it

provides higher performance on large datasets and

complex relationships of medical data, the Light

Gradient Boosting Machine (LightGBM), a fast

gradient boosting library, has been widely used. Some

authors used hybrid models, which were a

D., K. K., S., D., G., G. and N., K.

Hybrid Machine Learning Model for Early Prediction of Coronary Artery Disease: Integrating LightGBM and Ensemble Techniques for Enhanced Accuracy.

DOI: 10.5220/0013932400004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 5, pages

525-530

ISBN: 978-989-758-777-1

525

combination of ensemble model method together and

other models used together with the LightGBM

model in order to improve the prediction of Coronary

Artery Disease risk.

Chowdary et al. We (3. B. V. Chowdary, et.al.,

2021) make a distributed high-speed LightGBM-

based model for heart disease prediction and

weqarried to ascertain its efficiency in tackling bulk

medical data with high predictive accuracy.

Additionally, Gera and Melingi optimized a machine

learning ensemble approach to improve the Predictive

quality and Robustness of CAD risk models with

hyperparameter tuning and advanced loss functions

like (FL) (S. Gera and S. B. Melingi, 2023). The

relevance of data preprocessing and feature

extraction for making ML models perform better was

emphasized in their work. Additionally, the Enhanced

Whale Optimization Algorithm (EWOA) was

proposed as a feature selection approach and

incorporated the most critical risk factors in CAD

prediction, both improving the classification score

(Lakshmi and R. Devi, 2023).

Hybrid models: Features extracted from different

types of data can also be combined. In their work,

Gagoriya and Khandelwal analyzed several hybrid

ML techniques and emphasized the benefits of the use

of multiple algorithms for more accurate disease

classification and prediction (M. Gagoriya and M. K.

Khandelwal, 2023). Gupta et al. have made a

comparison between multiple classification models

like Decision Tree, Naïve Bayes, Random Forest, and

Logistic Regression in order to find the best approach

for the prediction of CAD (Gupta,et.al., 2023). They

found that ensemble techniques were superior to

standalone classifiers for predictive accuracy. Sharma

and Goel have further experimented with other

pathological approaches in ML. They had employed

SVM classification to find that AI-based predictive

models perform better compared to classical risk

score systems (R. Sharma and A. K. Goel, 2023).

XGBoost has also been widely used as an

essential algorithm in predicting CAD. Soni et al.

utilized XGBoost in building an effective predictive

model based on biomedical monitoring and wearable

device data, enhancing risk factor detection and early

diagnosis further (T. Soni,et.al., 2024). Their work

brought out the idea of bringing together ML and

health monitoring technology with real-time systems

to further the management of cardiovascular diseases.

D. P. K et al. also discussed the use of supervised and

unsupervised ML algorithms in detecting

cardiovascular diseases. According to the study, AI-

based models

increase

the

precision

myocytic

condition diagnosis, which occurs through predictive

analytics and feature selection techniques (S. Katari,

et.al., 2023).

Despite advancements, there is still potential to

enhance ML- based CAD prediction. Higher accuracy

and clinical relevance rely on ensemble methods,

hyperparameter tuning, and optimized loss functions.

This study proposes a hybrid ML model integrating

LightGBM with ensemble techniques to improve

CAD risk prediction and early detection.

3 MATERIALS AND METHODS

3.1 Dataset

The CAD dataset, consisting of 5,240 records from

the Framingham Heart Institute, was used to validate

the model. It includes various attributes such as sex

(Male = 1, Female = 0), age (continuous), smoking

status (1 = Yes, 0 = No), and the average number of

cigarettes smoked per day. Additional features include

the use of antihypertension drugs (1 = Yes, 0 = No),

history of stroke (1 = Yes, 0 = No), high blood

pressure (1 = Yes, 0 = No), and diabetes (1 = Yes, 0

= No). The dataset additionally includes overall

cholesterol levels, both systolic and diastolic blood

pressure, body mass index, heart rate, and blood

glucose levels. The target variable, Ten Year CAD,

indicates whether a patient developed coronary artery

disease within ten years (1 = Yes, 0 = No). A total of

15.19% of the records corresponded to patients with

CHD (744 cases), while 84.81% represented normal

cases (4,596 cases). Among the CHD patients,

53.26% were men, and 46.74% were women.

3.2 Data Cleaning and Preparation

Data cleaning and preparation plays a crucial role in

machine learning by ensuring that the data is high-

quality, dependable, and consistent prior to training

predictive models. Raw medical data are prone to

missing values, outliers, and imbalanced

distributions, which negatively impact model

performance. Therefore, systematic data

preprocessing techniques were used to address these

problems. Missing values were dealt with in the initial

step through the deletion of incomplete records or

imputing missing values with statistical imputation.

The presence of outliers was detected and deleted

using the IQR technique.

𝑰𝑸𝑹 = 𝑸𝟑 − 𝑸𝟏 (1)

where Q1 (lower quartile) and Q3 (upper quartile)

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

526

correspond to the 25th and 75th percentiles of the

data, respectively. The lower and upper boundaries

for outlier detection were determined using:

𝑩

𝒍

= 𝑸𝟏 − 𝟏.𝟓 ∗ 𝑰𝑸𝑹 (2)

𝐵

𝑢

= 𝑄3 + 1.5 ∗ 𝐼𝑄𝑅 (3)

3.3 Any Values beyond Bl and Bu

Were Considered Outliers and

Removed to Improve Model

Robustness: Evaluation Metrics

To quantitatively assess the performance of the hybrid

Light GBM-based model in predicting Coronary

Artery Disease, the metrics utilized include F1-score,

accuracy, precision, and recall. They evaluate the

predictive capacity of both positive and negative

instances. Specificity and sensitivity are extremely

useful for generating accurate patient prediction and

avoiding false positives. The Area Under the Curve

(AUC-ROC) is employed to measure the ratio of false

positives to true positives, reflecting the model's

ability to discriminate effectively. The confusion

matrix supplies the model error classification data for

optimization. Cross- validation methods dictate

generalizability and stability by avoiding overfitting.

The aforementioned measures collectively represent

comprehensive assessment of model predictive

ability making it ideal to be deployed at the clinical

site.

3.4 LightGBM

LightGBM is a scalability, speed, and efficiency-

optimized gradient boosting library that performs

well with large and high-dimensional data.

Differently from other traditional boosting

algorithms, LightGBM applies a leaf-wise policy of

tree growing that expands the leaf node with the

minimum loss, achieving rapid convergence and high

accuracy at low computational cost. LightGBM

supports categorical features naturally with minimal

preprocessing required (Abhishek, et.al ,2023).

In comparison with XGBoost, the former employs

a level- wise tree construction strategy and is

computationally expensive and slow when dealing

with enormously large datasets although it possesses

a robust regularization that is effective with most

tasks (S. Katari, et.al., 2023). CatBoost effectively

handles categorical variables through an ordered

boosting strategy with no target leakage, albeit with

the price tag of longer training time.

LightGBM is characterized by its training and

prediction speed, best missing value treatment, low

memory consumption, and good pattern recognition,

and it is thus among the top contenders for machine

learning applications needing speed and precision. Its

innovative design makes it particularly effective for

large-scale data analysis and real- time prediction

tasks. Figure 1 shows the Data Binning in LightGBM

For Faster Training.

Figure 1: Data binning in lightGBM for faster training.

3.5 Ensemble Techniques

Ensemble methods improve predictive power by

aggregating models. Bagging, boosting, and stacking

are common ensemble approaches. Boosting, adopted

in LightGBM, improves weak models iteratively by

targeting the misclassified instances. Stacking

aggregates heterogeneous models to capture their

strengths and ensure robustness. These methods

improve model generalization, and therefore they are

best applied in complex medical predictions.

LightGBM's efficiency, coupled with ensemble

techniques, makes it an ideal choice for medical

prediction tasks, where both speed and accuracy are

crucial.

4 EXPERIMENTAL RESULTS

Experimental results confirm that the LightGBM-

based model has a good predictive performance for

coronary heart disease. It promised high precision,

strong recall, accuracy, and F1-score for accurate

categorization. Compared to XGBoost and CatBoost,

LightGBM training was faster and more memory-

sparing while maintaining good predictive ability.

Furthermore, this SMOTE, we applied it also

addressed the class imbalance, thus increasing the

Hybrid Machine Learning Model for Early Prediction of Coronary Artery Disease: Integrating LightGBM and Ensemble Techniques for

Enhanced Accuracy

527

minority class recall. The model shows potential for

early disease detection, as tuned hyperparameters,

feature selection, and preprocessing, optimized the

model for robustness coronary artery disease

prediction LGBM roc curve the curve registers high

recall with relatively low false positive rate,

characteristic of great classification performance of

the model having AUC score of 97.8% The blue

curve shows the model's ability to tell classes apart,

which is a lot better than the red diagonal baseline that

shows random classification. The curve’s smooth and

sharp ascent indicates that the model successfully

detects positive cases with few misclassifications.

These findings affirm the model’s dependability for

precise disease prediction. (M. Gagoriya and M. K.

Khandelwal, 2023). Figure 2 shows the Overall

Model Performance.

Figure 2: Overall model performance.

Table 1: Evaluation of model performance metrics for

predicting coronary artery disease.

Models Auc Score Accuracy

Hybrid LightGBM

(Ensemble

Techni

ues

)

0.97 0.94

Light GBM (Single

Model)

0.95 0.90

XGBoost 0.91 0.89

CatBoost 0.92 0.90

Random Forest 0.88 0.87

The table 1 provides a thorough comparison of the

performance metrics across all models. The Hybrid

LightGBM (Ensemble Techniques) not only exceeds

the performance of the others in AUC score but also

achieves the highest accuracy, demonstrating its

strong effectiveness in CAD prediction. The

performance is also hierarchically better, with the

traditional models like Random Forest and AdaBoost

lagging behind, blessing the supremacy of employing

hybrid and ensemble approaches in machine learning

models when predicting health (S. Katari, et.al.,

2023). Figure 3 shows the Accuracy Over Training

Iterations for the Hybrid LightGBM Model.

Figure 3: Accuracy over training iterations for the hybrid

lightgbm model.

Moreover, the accuracy plot over training

iterations demonstrates the hybrid model capacity to

continue learning since accuracy improves and

reaches a peak of approximately 94% as training

iterations progressed. The progress indicates that the

model was developing and gaining knowledge from

the training data. These positive results suggest that

the hybrid LightGBM model l has the potential to be

an effective clinical decision support instrument by

providing accurate, prompt predictions of coronary

artery disease (CAD), which secures early treatment,

saves lives, and enhances the treatment of patients.

Future research will investigate incorporating

additional data types to increase predictive power,

i.e., lifestyle data or, genetic data etc. Furthermore,

real clinical trials validating the performance of the

model would verify its clinical utility and indicate

avenues for clinical improvement. In conclusion,

the results from our experiments further confirm the

acceptable application of the hybrid LightGBM

model for CAD prediction in the early clinical

diagnosis, highlighting its merits of accuracy,

efficiency and clinical applicability and offering a

promising candidate for changing the clinical

paradigm of CVD prediction and treatment.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

528

Table 2: Feature importance scores for coronary artery

disease prediction.

Feature Importance Score

Age 0.25

Cholesterol Level 0.20

Blood Pressure 0.15

Diabetes History 0.15

Smoking Status 0.10

Family History of CAD 0.10

BMI 0.05

These results suggest that the hybrid LightGBM

model is capable of identifying and assigning labels

to important risk factors, improving its predictive

performance and providing informative decision-

making advice in clinical practice. These

characteristics may help physicians to devise

preventive measures and approaches in CAD risk

patients. The experimental results validate the clinical

value of hybrid LightGBM model in early CAD

prediction, it can reflect the accuracy and

performance. This model performs well, therefore, its

use is suitable for early screening and serves as a

useful guide for improving patient outcomes. Future

research will attempt to improve its predictions and

apply it in clinics by augmenting the model with other

types of information, such as lifestyle or genetics.

Table 2 shows the Feature Importance Scores for

Coronary Artery Disease Prediction.

5 CONCLUSIONS

The manuscript presents HY OptGBM-Ensemble, a

hybrid machine learning model that combines

LightGBM with ensemble learning, Optuna-based

hyperparameter tuning, focal loss, and recursive

feature elimination. Using traditional models can

have a remarkable influence on the accuracy of the

prediction therefore this sort of methodology

improves the performance in tackling class imbalance

problem, feature selection problems and overall

prediction of the CGT quite significantly. HY

OptGBM- Ensemble gives better AUC (97.8%) than

Logistic Regression, Random Forest, and XGBoost

and even the individual LightGBM models. Often,

this leads to the generation of explainable AI models

that can be a reliable clinical decision support tool due

to improved interpretability, aided by SHAP analysis.

Future work that integrates deep learning capabilities

for low-complexity and real-time patient monitoring

and clinical decision notification may expand on the

initial successes with clinical data processing.

6 FUTURE WORK

Future studies may focus on optimization of a model

through the application of methods such as

convolutional neural networks (CNNs) or recurrent

neural networks (RNNs), which can help identify

complex underlying patterns in the healthcare data.

Real-time health care monitoring by wearable- based

and IoT- based health monitoring is possible in

improving prediction ability and early diagnosis. In

addition, there can be steps taken in increasing the

dataset size, i.e., inclusion of larger number of

populations and making use of fusion of multi-modal

data, i.e., genetics and imaging, to further improve

model generalizability. Further improvements can be

likely with more optimal fine-tuning of feature

selection techniques and with research into

interpretability of the AI methods used thereby

increasing the trust and acceptance of clinical end-

users.

REFERENCES

A. Lakshmi and R. Devi, "Feature Selection for Heart

Disease Prediction Using an Enhanced Whale

Optimization Algorithm," 2023 12th International

Conference on System Modeling & Advancement in

Research Trends (SMART), Moradabad, India, 2023,

pp. 644- 648. doi: 10.1109/SMART59791.2023.1042

8617.

A. Gupta, V. Misra, K. Chauhan, and K. Manoj,

"Leveraging Machine Learning for Heart Disease

Prediction," 2023 5th International Conference on

Advances in Computing, Communication Control and

Networking (ICAC3N), Greater Noida, India, 2023, pp.

108- 112. doi: 10.1109/ICAC3N60023.2023.10541622

Abhishek, H. V. Bhagat, and M. Singh, "An ML-Based

Approach for Early Detection of Cardiovascular

Diseases," 2023 Second International Conference on

Advances in Computational Intelligence and

Communication (ICACIC), Puducherry, India, 2023,

pp. 1-5. doi: 10.1109/ICACIC59454.2023.10435210.

Hybrid Machine Learning Model for Early Prediction of Coronary Artery Disease: Integrating LightGBM and Ensemble Techniques for

Enhanced Accuracy

529

Anurag, S. Kalta, J. Thakur, H. Bhardwaj, and Y. Banyal,

"An Optimized Ensemble Machine Learning Model for

Heart Disease Prediction," 2023 3rd Asian Conference

on Innovation in Technology (ASIANCON), Ravet,

India, 2023, pp. 1- 5.doi: 10.1109/ASIANCON58793.

2023.10270436

B. V. Chowdary, J. Kedarnath, R. Vyshnavi, V. Lavakush,

and C. Shashidhar, "A Distributed Prediction Model for

Heart Disease Using LightGBM," 2021 Fifth

International Conference on I-SMAC (IoT in Social,

Mobile, Analytics, and Cloud) (I-SMAC), Palladam,

India, 2021, pp. 662-667. doi: 10.1109/I-

SMAC52330.2021.9640897.

B. N. Shirazi, S. Naghsh, A. A. Safavi, and A.

Sharafkhaneh, "Utilizing Machine Learning for

Coronary Heart Disease Prediction," 2023 9th

International Conference on Control, Instrumentation

and Automation (ICCIA), Tehran, Iran, 2023, pp. 1-5.

doi: 10.1109/ICCIA61416.2023.10506325.

M. Gagoriya and M. K. Khandelwal, "Hybrid Approach in

Machine Learning for Analyzing Heart Disease," 2023

International Conference on Intelligent and Innovative

Technologies in Computing, Electrical and Electronics

(IITCEE), Bengaluru, India, 2023, pp. 896-899. doi:

10.1109/IITCEE57236.2023.10090896.

M. T. R, M. Mohamed Musthafa, A. M. J. M. Zubair

Rahman, J. Viswanath, K. Vanitha, and V. K. V,

"Ensemble- Based Heart Disease Prediction with

Reduced Entropy," 2024 IEEE International

Conference on Interdisciplinary Approaches in

Technology and Management for Social Innovation

(IATMSI), Gwalior, India, 2024, pp. 1-5. doi:

10.1109/IATMSI60426.2024.10503147.

R. Sharma and A. K. Goel, "Machine Learning

Applications for Diagnosing Heart Disease," 2023 3rd

International Conference on Technological

Advancements in Computational Sciences (ICTACS),

Tashkent, Uzbekistan, 2023, pp. 1521-1525. doi:

10.1109/ICTACS59847.2023.10390206.

S. Katari, T. Likith, M. P. S. Sree, and V. Rachapudi,

"Hybridized ML Algorithms for Predicting Heart

Disease," 2023 International Conference on Sustainable

Computing and Data Communication Systems

(ICSCDS), Erode, India, 2023, pp.

121125.doi:10.1109/ICSCDS56580.2023.10104609.

S. Gera and S. B. Melingi, "Optimized Ensemble Learning

for Predicting Heart Disease," 2024 Third International

Conference on Electrical, Electronics, Information, and

Communication Technologies (ICEEICT),

Trichirappalli, India, 2024, pp. 1-8. doi:

10.1109/ICEEICT61591.2024.10718534.

T. Soni, D. Gupta, and M. Uppal, "XGBoost-Based Risk

Factor Analysis for Cardiovascular Disease

Prediction," 2024 IEEE 3rd World Conference on

Applied Intelligence and Computing (AIC), Gwalior,

India, 2024, pp. 353-358. doi:

10.1109/AIC61668.2024.10730877.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

530