Hybrid Gradient Boosting and Graph Neural Network Architecture

for Cardiovascular Disease Prediction

Kummetha Snehalatha, Mudhavarthi Phabe Zoe Gospel, Y. Indira Priyadarshini,

Pasam Hima Teja and Kethapally Sri Sathwika

Department of CSE, Ravindra College of Engineering for Women, Kurnool, Andhra Pradesh, India

Keywords: Cardiovascular Disease Prediction, Graph Neural Networks, Gradient Boosting, Models, Hybrid Machine

Learning, Graph‑Structured Data.

Abstract: Heart disease (CVD) is one of the main causes of death around the world, so an advanced future model is

required to assess sufficient risk. Although traditional machine learning techniques such as Gradient's

Boosting Machines (GBMS) are effective with structured spot data, they cannot describe complex graph -

based conditions usually found in health data. Conversely, Graph Neural Networks (GNNs) are suitable for

graph-structured data but not tabular data. To fill this gap, this research proposes a hybrid model that combines

GBM and GNN for CVD prediction, leveraging the strengths of both methods. Using the Framingham Heart

Study dataset, a synthetic graph is built to capture patient similarities. Experimental results indicate that the

proposed hybrid method outperforms individual models in accuracy and AUC-ROC, proving its ability to

improve predictive performance and inform healthcare analytics using multimodal data fusion.

1 INTRODUCTION

1.1 Background

Cardiovascular diseases (CVDs)(M. Močnik and N.

Marčun Varda, 2024) account for about 32% of

annual deaths worldwide, making early detection and

prevention ever more necessary. Heart attacks and

strokes tend to result from an interaction of hereditary

factors, lifestyle, and environmental factors. Machine

learning (ML) algorithms have been reliable

predictors of CVD risk (V. V. Paul and J. A. I. S.

Masood , 2024) by identifying intricate patterns in

large amounts of data in recent years. Of these,

Gradient Boosting Models (GBMs) such as

LightGBM and XGBoost perform very well in

handling tabular data by best capturing complex

feature interactions with a minimal chance of

overfitting when well-optimized.

But GBMs are unable to model internal patterns

of health data, such as demographic consent, general

medical history or lifestyle habits in patients. These

inter oriented patterns are more effectively depicted

with graph -based models. Graff Neural Networks

(GNNS), engineers for graph -composed data,

expanding the classic nervous network with a

message -focused mechanism to integrate

relationship dependence. GNN has important

applications in the health care system(B. Khemani et

al., 2024). where patients are an important factor in

predicting the disease, which includes referral

relationships, co-ligging or health social

determinants, predict the disease and assess the risk.

1.2 Problem Statement

While current CVD prediction models are effective

on tabular data, they tend to ignore the rich

information contained in graph-based relationships

within the data. This drawback hinders their potential

to capture meaningful patient-to-patient

interactions(Y. Liang, et al., 2024) that can improve

predictive performance. (R. J. Woodman and A. A.

Mangoni, 2023) The absence of a hybrid model that

connects both data correctly -makes a difference in

existing CVD prediction methods. To fill this

difference, an integrated approach is required - to

promote the risk of disease risk to use GNN's -related

modeling power and the forecasting power of GBMS.

Snehalatha, K., Gospel, M. P. Z., Priyadarshini, Y. I., Teja, P. H. and Sathwika, K. S.

Hybrid Gradient Boosting and Graph Neural Network Architecture for Cardiovascular Disease Prediction.

DOI: 10.5220/0013881600004919

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 2, pages

287-291

ISBN: 978-989-758-777-1

287

1.3 Contribution

This work addresses the challenges identified above

by proposing a new hybrid model that combines

GNNs and GBMs to enhance CVD prediction. The

approach has three contributions. It builds a synthetic

graph-based patient similarity network with edges

corresponding to demographic and clinical feature

similarities. Second, the GNN module produces

graph embeddings capturing these relationships,

which are then combined with the initial tabular

features to be used as inputs to the GBM. Finally, an

extensive experimental evaluation is performed,

showing that the hybrid model outperforms single-

facet methods in precision, AUC-ROC, and other

metrics of performance. The study shows the promise

of combining more than one representation of the data

in order to improve predictive performance in health

applications (H. O. Boll, 2025).

2 LITERATURE REVIEW

2.1 Graph Neural Networks in

Healthcare

Recent progress in Graph Neural Networks (GNNS)

has inspired their use in various health applications(S.

G. Paul,et.al., 2024) such as Genomics, Drug

Discovery, and Patient Network Analysis. In

genomics, GNN is used to represent the gene

regulator network, to detect complex relationships

between genes. Cases of drug discovery also use

GNNs, which are for predictions by handling

chemical structures in the form of graphs on

molecular properties and interactions. Inpatient

network analysis, GNNS is best suited for learning-

related patterns, including referral networks, disease-

cum phenomenal ratings, or the relationship between

social determinants for health. While being flexible,

the need for GNN for graph-composed data limits

them from being independent use, especially in

contexts where data sets are mainly present in a table

format. This aspect highlights the need for coupling

with models that work well with structured data.

2.2 Gradient Boosting Models

Gradient Boosting Models (GBMs) have been a

workholder in the health care analysis which is a

future health analysis due to their performance and

efficiency on structured data. (T. V. Afanasieva,et.al.,

2023)and health studies, which process health

services for their ability to treat, and handle lack of

values, handle lack of values , and models of non-led

interactions. In the prediction of heart disease, GBM

is regularly used for data such as the Framingham

Heart Study, where inputs such as age, cholesterol,

and smoking conditions are used to predict exits such

as 10 years risk of coronary heart disease.

Unfortunately, GBM has no means to integrate

relationships between classes or relevant knowledge,

which can promote the use of their future in more

complex situations.

2.3 Hybrid Approaches

Hybrid Machine Learning Framework has recently

begun to gain popularity as a possible option to

include many data tours. Combined nerve network

researchers with traditional models such as GBM can

enable researchers to benefit from the properties of

both streams. For example, hybrid models, customer

partitions, fraud, and multi-models have proved

valuable in applications related to health analysis.

Nevertheless, there are smaller tasks that study the

combination of GNN and GBMs to get the graph

structure as well as table data at the same time. This

lack of literature provides an important opportunity to

improve the future model through the rich

representation of the prediction of heart disease(H. A.

Al-Alshaikh et al., 2024), to increase the accuracy

and strength of tasks such as prediction of heart

disease.

3 METHODOLOGY

3.1 Dataset

Framingham Heart Study data provides the basis for

this study, which provides a complete range of

demographic, behavior and clinical properties

associated with heart disease. The age of special

significance, gender, cholesterol levels, systolic and

diastolic blood pressure, diabetes and smoking. The

target variable, "Tenyarchd," is an indicator of

whether coronary heart disease exists within 10 years.

The balanced combination of a categorized and

constant variable in the dataset makes it suitable for

both tabulated and graph -based analysis.

3.2 Graph Construction

A synthetic graph has been created to represent the

relationship between patients based on gender

equality. Knots in the graph are individual patients,

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

288

and the edges are made between the nodes if the

distance to the Euclidian between their functional

vectors is lower than a given threshold. The weight of

the edges is used to represent the level of equality,

which allows GNN to learn a relationship. The effect

of this graph -based(A. Shraga and B. Or, 2024)

construction method is that patients with analog

health profiles are more closely linked, causing

meaningful associations of the model.

3.3 Hybrid Model Architecture

The proposed hybrid structure consists of two

components:

• Graph Neural Network (GNN): The Two-

Team Graph Conversion Network (GCN)

calculates the graph created to produce node

entry. The patient-specific interactions

achieved from the node that involves the

graph capture the relationship.

• Grade-GBM: Lightgbm, a highly adapted

GBM implementation, is used to handle

basic functions, such as original table data

and GNN-Janite built-in. This merger

benefits from the strength of both models so

that hybrid architecture can effectively

manage different data representations.

4 EXPERIMENTAL DESIGN

4.1 Data Preprocessing

To ensure the best performance of the hybrid model,

the following preprocessing is done:

• Missing Value Handling: Average modeling

is used to change the missing values in

numeric plants.

• Normalization: Numerical functions are

generally normalized to make compatible

with the GNN component.

• Categorical Encoding: The area is coded

through A-warm coding to feed the cracked

variables in the GBM component.

4.2 Training

• GNN Training: The GCN team is trained on

cross-raising losses for learning rates and

node classification of 0.01 with Adam

Optimizer.

• GBM Training: LightGBM is trained with

standard parameters, such as a learning

speed of 0.1 and maximum depth of 6. Initial

limitations to avoid overfit.

4.3 Evaluation Metrics

The hybrid model is assessed based on the following

performance metrics:

• Accuracy: Measures the overall correctness

of predictions.

• AUC-ROC: It measures how well the model

can differentiate between positive and

negative instances.

• Precision and Recall: Evaluate the model’s

performance on correctly identifying true

positives and avoiding false negatives.

5 RESULTS

5.1 Performance Metrics

Comparison of the GNN-only, GBM-only, and

Hybrid Model in Figure 1 brings forth the advantages

of combining graph-based learning with the

conventional machine learning methods. The only

GNN model achieved 0.78 accuracy and AUC-RC of

0.81, which clearly shows the potential of the

interactive structural extraction, but weakened with

table data. On the other hand, the GBM-only model

reached better, reached a measure of the accuracy of

0.82 and a measure of AUC-RC of 0.84, better in

structured data analysis without any relationship with

any relationship. The hybrid model, which uses a

combination of both methods, is performed better

than individual models of 0.88 accuracy and AUC-

Roc of 0.91, which reflects the benefits that come

with the integration of graph (Figure 2) that is built

into traditional machine learning.

These results suggest that integrating GNN

embeddings into GBM can significantly enhance

predictive performance, making it a valuable

approach for cardiovascular disease prediction and

broader healthcare applications. The details are

tabulated at Table1.

Table 1: Comparison Between GNN, GBM and Hybrid

Model.

Model Accuracy AUC-ROC

GNN Only 0.78 0.81

GBM Only 0.82 0.84

brid Model 0.88 0.91

Hybrid Gradient Boosting and Graph Neural Network Architecture for Cardiovascular Disease Prediction

289

Figure 1: Comparison Between GNN, GBM and Hybrid Model.

5.2 Visualization

• ROC Curves: The hybrid model achieves a

significantly higher AUC compared to the

standalone GNN and GBM models,

illustrating its superior predictive capability.

Figure 2: ROC Curve with True Positive Rate and False

Positive Rate.

• t-SNE Embeddings: Visualization of the

GNN-generated embeddings reveals distinct

clusters corresponding to different patient

profiles, highlighting the model’s ability to

capture relational nuances.

6 DISCUSSION

6.1 Insights

Increased performance of hybrid models highlights

the benefits of integrating graph-based and table data

representation. The GNN module captures the patient

relationship effectively, which complements GBM's

stronger ability to handle structured functions. This

integration allows hybrid architecture to make more

accurate and explanatory predictions.

6.2 Limitations

Synthetic graph construction: Dependency on the

construction of synthetic graph can limit the use of

models in practical applications where graph

structures are not so well defined.

Data cost: Training for hybrid models requires

adequate calculation power, especially in the case of

large data.

7 FUTURE WORK

Further in the future, it will aim to broaden the use of

hybrid frameworks for well-defined graph structures,

like electronic health records and actual health

services datasets that include the patient's referral

network. To promote the representation of the

patient's conditions, researchers can fix the model

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

290

architecture by using innovative (W. Yang et al.,

2024) Adapt calculation efficiency using techniques

such as pruning and distributed training will allow

scalability on large datasets. Integration of

clarification methods will improve the interpretation

of the model, which will lead to high faith in the

clinical environment (J. A. Damen et al. ,2016).

Finally, the extension of frameworks for many data

sources such as genomic and imaging data can unlock

new applications for accurate therapy and early

detection of diseases.

8 CONCLUSIONS

The contribution from this study is a demonstration

of remarkable advantage of integrating Graph Neural

Networks (GNN) and Grade Bosting Machines

(GBM) against the prediction of heart disease (CVD).

The hybrid model originally uses GNN to highlight

complex conditions from graph -composed data and

uses GBM's ability to handle table data. Through the

merger of these orthogonal approaches, the proposed

structure crossed individual models, and showed

remarkable performance in the accuracy of the

prediction and AUC-ROC performance.

In addition to its future indicative capacity, this

method highlights the widespread prevention of

hybrid modeling in healthcare analysis, especially in

an environment where both spot and relationship data

exist. The results suggest that the use of the equality

network can emphasize the latent patterns, and

provide more insight into the development of the

disease and related risk factors. Such methods can

improve initial diagnosis and enable targeted

interventions, which require reducing CVD-related

mortality. Nevertheless, there are challenges, such as

the use of artificial graph structures and high

calculation requirements. Future work should be

aimed at using this structure on real datasets and

scaling it. Overcoming these challenges will enable

the hybrid model to become a valuable tool for

accurate therapy, which continues personal health

services.

REFERENCES

A. Shraga and B. Or, “Explainable Prediction of

Cholesterol Levels in Women Using Decision Tree

Models,” Authorea Preprints, Accessed:

Mar. 13, 2025. [Online]. Available: https://www.techr

xiv.org/doi/full/10.36227/techrxiv.173398176.658954

B. Khemani et al., “Sentimatrix: sentiment analysis using

GNN in healthcare,” Int. j. inf. tecnol., vol. 16, no. 8,

pp. 5213–5219, Dec. 2024, doi: 10.1007/s41870-024-

02142-z.

H. O. Boll, “Graph neural networks for clinical risk

prediction based on patient similarity graphs,” 2024,

Accessed: Mar. 12, 2025. [Online]. Available:

https://lume.ufrgs.br/handle/10183/283321

H. A. Al-Alshaikh et al., “Comprehensive evaluation and

performance analysis of machine learning in heart

disease prediction,” Scientific Reports, vol. 14, no. 1, p.

7819, 2024.

J. A. Damen et al., “Prediction models for cardiovascular

disease risk in the general population: systematic

review,” bmj, vol. 353, 2016, Accessed: Mar. 13, 2025.

[Online].

Available:https://www.bmj.com/content/353/bmj.i241

6.abstract

M. Močnik and N. Marčun Varda, “Preventive

Cardiovascular Measures in Children with Elevated

Blood Pressure,” Life, vol. 14, no. 8, p. 1001, Aug.

2024, doi: 10.3390/life14081001.

R. J. Woodman and A. A. Mangoni, “A comprehensive

review of machine learning algorithms and their

application in geriatric medicine: present and future,”

Aging Clin Exp Res, vol. 35, no. 11, pp. 2363–2397,

Sep. 2023, doi: 10.1007/s40520-023-02552-2.

S. G. Paul, A. Saha, M. Z. Hasan, S. R. H. Noori, and A.

Moustafa, “A systematic review of graph neural

network in healthcare-based applications: Recent

advances, trends, and future directions,” IEEE Access,

vol. 12, pp. 15145–15170, 2024.

T. V. Afanasieva, A. P. Kuzlyakin, and A. V. Komolov,

“Study on the effectiveness of AutoML in detecting

cardiovascular disease,” Aug. 19,

2023, arXiv: arXiv:2308.09947. doi: 10.48550/arXiv.2

308.09947.

V. V. Paul and J. A. I. S. Masood, “Exploring predictive

methods for Cardiovascular Disease: a survey of

methods and applications,” IEEE Access, 2024,

Accessed: Mar. 12, 2025. [Online].

Available:https://ieeexplore.ieee.org/abstract/documen

t/10606182/

W. Yang et al., “Deciphering cell– cell communication at

single-cell resolution for spatial transcriptomics with

subgraph-based graph attention network,” Nature

Communications, vol. 15, no. 1, p. 7101, 2024.

Y. Liang, C. Guo, and H. Li, “Comorbidity progression

analysis: patient stratification and comorbidity

prediction using temporal comorbidity network,”

Health Inf Sci Syst, vol. 12, no. 1, p. 48, Sep. 2024, doi:

10.1007/s13755-024-00307-5.

Hybrid Gradient Boosting and Graph Neural Network Architecture for Cardiovascular Disease Prediction

291