Heart Attack Risk Detection Using Eye Retinal Images
Heart Attack Risk Detection Using Eye Retinal Images
Keywords: Retinal Imaging, Cardiovascular Diseases, Heart Attack Prediction, Artificial Intelligence, Non-Invasive Diagnostics,
Recurrent Neural Networks, Expectation-Maximization.
Rajani; Paavana Guruprasad; Rahul M Javkar; Rakshith C R; Rakshitha T S, (2025), Heart Attack Risk Detection
Using Eye Retinal Images. International Journal of Innovative Science and Research Technology,
10(1), 756-761. https://doi.org/10.5281/zenodo.14730631
756
Heart Disease Prediction Using Eye Retinal Images indicator of vascular health and demonstrated UNET's
(Shaikh, S. et al., 2023): effectiveness in accurate vessel segmentation. Challenges
This study employed transfer learning with the Inception included the reliance on datasets like MESSIDOR and
v3 model and convolutional neural networks (CNNs) to DRIVE, which limited generalizability, and the "black-box"
analyze over 6000 retinal images, achieving impressive nature of VGG-16, which created interpretability issues
accuracy rates of 97% during training and 96% during critical for clinical adoption. The methodology involved
testing. The research demonstrated the effectiveness of preprocessing images from datasets, segmenting retinal
automated feature extraction for detecting cardiovascular vessels using UNET, and computing AVR for risk
risks and highlighted the value of transfer learning in assessment. This non-invasive diagnostic approach offers
addressing challenges associated with smaller datasets. By high accuracy, automated segmentation, and early detection
utilizing a robust yet efficient model, the study paved the of cardiovascular risks, further strengthening its potential for
way for non-invasive diagnostic methods that require clinical use. Expanding dataset diversity and enhancing
minimal manual intervention. However, limited diversity in model explainability are critical steps for future research in
the dataset affected the model's generalizability, and its this domain.
reliance on specialized imaging equipment, along with the
opaque "black-box" nature of CNNs, hindered clinical Heart Attack Risk Prediction Using Retinal Images
adoption. The methodology involved preprocessing retinal (Pailla Teena Reddy et al., 2023):
images, applying transfer learning with Inception v3 for This study focused on extracting features from retinal
feature extraction, and validating the model's performance. images to detect patterns indicative of heart attack risks,
This non-invasive diagnostic approach offers early detection employing supervised learning methods for prediction. It
capabilities, high accuracy, and automated feature emphasized the potential of retinal imaging as a non-
extraction, which collectively enhance its potential for invasive and scalable diagnostic tool for early detection and
improving cardiovascular risk prediction. Future research population-level screening. By bridging the gap between
could explore integrating explainable AI methods to data accessibility and early-stage diagnosis, this study
improve the interpretability of the results for clinical highlighted the advantages of automated techniques in
acceptance. resource-limited environments. However, challenges
included the limited availability of advanced imaging
Cardiovascular Disease Prediction from Retinal Images systems in resource-constrained regions and the need for
Using Machine Learning (Rose, B. et al., 2023): extensive validation across diverse datasets to ensure
This study explored the use of Contrast Limited clinical reliability. The methodology involved preprocessing
Adaptive Histogram Equalization (CLAHE) to enhance retinal images to enhance relevant features, using algorithms
retinal vessel segmentation, comparing the performance of for feature extraction, and training supervised learning
Support Vector Machines (SVM) and CNNs. SVM achieved models for risk prediction. The system provides early
superior accuracy of 97.46%, emphasizing its efficiency detection capabilities, automated analysis, and accessibility,
over CNN in this context. The study demonstrated the making it a safer and more efficient alternative to traditional
significant role of preprocessing techniques like CLAHE in diagnostic methods. Future work could focus on integrating
boosting the clarity of critical retinal features, improving cloud-based solutions to make this approach widely
overall model performance. The research underscored the accessible in low-resource settings.
importance of preprocessing techniques like CLAHE in
improving segmentation and identifying retinal features such Prediction of Cardiovascular Risk Factors from Retinal
as vessel diameter and tortuosity, which are linked to Fundus Images via Deep Learning (Poplin, R. et al.,
cardiovascular health. Challenges included computationally 2018):
intensive feature extraction, limiting scalability, and the use This study utilized convolutional neural networks
of small, non-representative datasets, which increased the (CNNs) trained on datasets from UK Biobank and EyePACS
risk of overfitting. The methodology involved CLAHE for to predict cardiovascular risk factors such as age, gender,
image contrast enhancement, feature extraction based on and smoking status from retinal fundus images. The research
vessel attributes, and classification using SVM and CNN for achieved high accuracy in predicting key cardiovascular
comparison. The study demonstrated the non-invasive indicators, demonstrating its potential for risk stratification
diagnostic potential of retinal imaging, with high accuracy and offering performance comparable to existing clinical
and enhanced segmentation capabilities using CLAHE. It risk calculators. By establishing correlations between retinal
also emphasized the need for developing lighter algorithms features and systemic risk factors, the study provided a
to improve scalability and deployment in real-world settings. novel avenue for non-invasive cardiovascular diagnostics.
Challenges included limited dataset diversity, which
Identifying Retinal Abnormalities for Cardiovascular restricted broader applicability, and the absence of
Risk Assessment (Prakash, A. P. et al., 2024): biochemical markers like lipid panels, which reduced the
This research introduced the Arteriolar to Venular comprehensiveness of predictions. The methodology
Ratio (AVR) as a biomarker for cardiovascular risk involved applying deep learning techniques to retinal fundus
assessment and utilized the UNET architecture for precise images, training CNN models on large datasets, and
retinal vessel segmentation. The classification model based evaluating the models using metrics like MAE and AUC.
on VGG-16 achieved an accuracy of 97%, showcasing its This non-invasive approach supports population-level
reliability. With its ability to detect subtle vascular screenings, providing efficient and accurate predictions of
irregularities, the study demonstrated the relevance of AVR cardiovascular risk factors. Scaling up this approach and
as an indicator for predicting systemic cardiovascular incorporating additional health indicators could further
conditions. The study highlighted the utility of AVR as an enhance its clinical relevance and reliability.
757
III. OBJECTIVES easy to use. The fig 3 shows the architecture of the proposed
system.
Achieve a highly accurate model that enables early
detection of heart attack risks, facilitating timely medical Data Collection:
intervention. Retinal images were sourced from the RFMiD dataset,
Build a user-friendly website where users can upload which provides a variety of images required for training and
retinal images and receive real-time risk assessments. testing the system.
Use a dataset representing a diverse population with
varying retinal patterns to ensure generalization and Preprocessing:
prevent bias. To improve the quality and consistency of the input
Develop a non-invasive heart attack risk detection data, the following techniques were used:
system using retinal imaging as a safer and more Image Augmentation: Enhanced the dataset by adding
accessible alternative to traditional diagnostic methods. variations through rotation, flipping, and other
Upload processed data and risk assessment results to a transformations.
cloud-based platform for real-time monitoring and Square Padding: Ensured that all images had the same
remote access by caregivers or healthcare professionals. dimensions by adding padding where needed.
Retinal Cropping: Focused only on the relevant parts of
IV. PROPOSED SYSTEM the retina, removing unnecessary details from the image.
Resizing: Standardized image sizes to make them
The proposed system is a non-invasive, AI-based compatible with the AI model.
solution designed to predict heart attack risks by analyzing Normalization: Scaled the pixel values to a uniform
retinal images. Traditional methods for heart attack risk range, improving the model's ability to process the
assessment often involve expensive and invasive images effectively.
procedures. This system takes a simpler and safer approach
by using the similarities between retinal and cardiovascular Feature Analysis and Clustering:
blood vessels to identify potential risks. Advanced deep
learning algorithms help detect patterns in the retinal images Clustering: Used Density-based clustering algorithm to
linked to heart health. group similar patterns or features in the retinal images
based on density of blood vessels, to make the data more
Key features of the Proposed System Include: organized for the next stage, as shown in fig 1.
Non-Invasive Approach:
It uses retinal images to detect heart attack risks early
without the need for invasive procedures.
Smart AI Models:
Combines Recurrent Neural Networks (RNN) with
Expectation-Maximization (EM) for analyzing patterns. This
is different from other systems that mostly use CNNs or
SVMs. The RNN helps pick up sequences and trends in the
data, making the predictions more accurate.
Better Preprocessing:
The system uses techniques like cropping, resizing, and FIG 1: IMAGE CLUSTERING
normalization to ensure the data is clean and consistent. It
also balances the dataset to avoid biased results, something Model Training:
many existing studies don’t focus on enough.
The system used Recurrent Neural Networks (RNN),
Easy-to-Use Web Platform: particularly LSTM and GRU, to detect patterns in the
Users can upload retinal images on a simple web data.
interface and get real-time analysis, making it practical and
These features were then classified using an Expectation-
accessible for both healthcare professionals and individuals.
Maximization (EM) model, combining the strengths of
both approaches.
Affordable and Scalable:
The training process was optimized using:
It is designed to be cost-effective and adaptable, with
careful consideration for different populations and varying Categorical Cross-Entropy as the loss function to
health conditions. measure prediction accuracy.
Adam Optimizer to adjust model weights and improve
performance.
V. IMPLEMENTATION
Batch Size, Epochs, and Validation were tuned for better
The implementation of the proposed system involved results.
several important steps, starting from preparing the data to
training and deploying the model. Each step was carefully
designed to ensure the system performs accurately and is
758
Evaluation: Precision ⋅ Recall
F1-Score = 2 ⋅
The model was tested using metrics such as: Precision + Recall
Accuracy: It measures the overall correctness of the
model by calculating the proportion of correctly
classified instances out of the total instances. It gives a The results confirmed the system’s ability to make
general measure of model performance but can be reliable predictions.
misleading in imbalanced datasets.
Deployment:
True Positives + True Negatives The final system was deployed on a web-based
Accuracy = platform, allowing users to upload retinal images and
Total Number of Samples
receive real-time analysis. The platform ensures
Precision: It measures the proportion of correctly accessibility and ease of use, making it practical for
predicted positive observations to the total predicted everyday healthcare applications.
positives
VI. RESULT AND ANALYSIS
True Positives
Precision = The heart attack risk detection system demonstrated
True Positives + False Positives
exceptional performance, achieving an accuracy of 98.6%
Recall (Sensitivity): It measures the proportion or actual with the available data set, which highlights its capability to
positives correctly identified by the model. It focuses on classify heart attack risks with high reliability.
how well the model captures positive cases out of all
actual positive cases. The term sensitivity also refers to
the true- positive rate.
True Positives
Recall =
True Positives + False Negatives
759
Fig 3: System Architecture
The project showcases how artificial intelligence can [1]. Shaikh, S.; Shaikh, M.; Khan, A., Patel, R.; Ali, Z.,
revolutionize healthcare by using retinal images to predict "Heart Disease Prediction Using Eye Retinal Images,"
heart attack risks. By incorporating advanced deep learning 2023 6th International Conference on Advances in
techniques like RNNs and clustering, the system is able to Science and Technology (ICAST), Mumbai, India,
detect even subtle vascular changes associated with heart 2023, pp. 417-422, doi:
health issues. Its non-invasive, easy-to-use, and affordable 10.1109/ICAST59062.2023.10454943.
design makes it particularly beneficial for increasing access [2]. Rose, B.; Smith, T.; Johnson, L., "Cardiovascular
to preventive healthcare, especially in areas with limited Disease Prediction from Retinal Images using Machine
medical resources. By providing reliable risk predictions and Learning," 2023 International Conference on
actionable health advice, the system overcomes many of the Sustainable Computing and Data Communication
challenges posed by traditional diagnostic methods, Systems (ICSCDS), Erode, India, 2023, pp. 1-5, doi:
promoting early detection and timely treatment. As the 10.1109/ICSCDS56580.2023.10104816.
system evolves, refining it to work effectively with diverse [3]. Prakash, Y.P.; Kumar, S.; Gupta, A.; Sharma, P.,
datasets and undergoing thorough clinical validation will be "Identifying the Abnormalities in Retinal Images
key to its widespread acceptance. This innovation has the Towards the Prediction of Cardiovascular Disease
potential to make a real difference in reducing deaths caused Using Deep Learning," 2024 International Conference
by heart disease and marks a step forward in using AI to on Wireless Communications Signal Processing and
create smarter healthcare solutions worldwide. Networking (WiSPNET), Chennai, India, 2024, pp. 1-
6, doi: 10.1109/WiSPNET61464.2024.10532859.
Future enhancements for the project focus on [4]. Reddy, P.T.; Ramesh, V.; Kumar, K.; Rao, M., "Heart
expanding the dataset by including more retinal images from Attack Risk Prediction Using Retinal Eye Images,"
diverse demographics, which will help improve the model's IJFANS International Journal of Food and Nutritional
ability to generalize across different populations. Sciences, vol. 12, no. 4, pp. 875-878, 2023.
Collaborating with healthcare institutions for clinical [5]. Poplin, R.; Varadarajan, A.V.; Blumer, K.; Liu, Y.;
validation will ensure the system meets regulatory standards McConnell, M.V.; Corrado, G.S.; Peng, L.; Webster,
and can be seamlessly integrated into real-world medical D.R., "Prediction of cardiovascular risk factors from
workflows. Additionally, the system could be extended to retinal fundus photographs via deep learning," Nature
assess risks for other cardiovascular conditions, further Biomedical Engineering, vol. 2, no. 3, pp. 158-164,
broadening its scope and impact. Optimizing the platform Mar. 2018, doi: 10.1038/s41551-018-0195-0.
for mobile devices would make this technology more [6]. Roth, G.A.; Mensah, G.A.; Johnson, C.O.; Addolorato,
accessible, especially in underserved and resource-limited G.; Ammirati, E.; Baddour, L.M.; Barengo, N.C.;
areas, enhancing its potential to improve public health on Beaton, A.Z.; Benjamin, E.J.; Benziger, C.P., "Global
a broader scale. burden of cardiovascular diseases and risk factors", J.
Am. Coll. Cardiol. Vol 76, pp.2982–3021, 2020.
760
[7]. Stone, N.J.; Robinson, J.; Lichtenstein, A.H.; Bairey
Merz, C.N.; Blum, C.B.; Eckel, R.H.; Goldberg, A.C.;
Gordon, D.; Levy, D.; Lloyd-Jones, D.M.; et al., "2013
ACC/AHA guideline on the treatment of blood
cholesterol to reduce atherosclerotic cardiovascular
risk in adults: A report of the American College of
Cardiology/American Heart Association Task Force on
Practice Guidelines," Circulation, vol. 129, pp. S1–
S45, 2014.
[8]. Goff, D.C.; Lloyd-Jones, D.M.; Bennett, G.; Coady, S.;
D’Agostino, R.B.; Gibbons, R.; Greenland, P.;
Lackland, D.T.; Levy, D.; O’Donnell, C.J.; et al.,
"2013 ACC/AHA guideline on the assessment of
cardiovascular risk: A report of the American College
of Cardiology/American Heart Association Task Force
on Practice Guidelines," Circulation, vol. 129, pp.
S49–S73, 2014.
[9]. Ting, D.S.W.; Cheung, C.Y.L.; Lim, G.; Tan, G.S.W.;
Quang, N.D.; Gan, A.; Hamzah, H.; Garcia-Franco, R.;
San Yeo, I.Y.; Lee, M.L.; Wong, T.Y., "Development
and validation of a deep learning system for diabetic
retinopathy and related eye diseases," JAMA, vol. 318,
pp. 2211–2223, 2017.
[10]. Cheung, C.Y.; Hsu, W.; Lee, M.L.; Wang, J.J.;
Mitchell, P.; Lau, Q.P.; Kawasaki, R.; Klein, R.; Klein,
B.E.; Cotch, M.F.; Wong, T.Y., "Retinal vascular
fractal dimension and its relationship with
cardiovascular and ocular risk factors," American
Journal of Ophthalmology, vol. 154, pp. 663–674.e1,
2012.
761