Retina-Cardiac Screening: Risk Prediction of Heart Attack using Retinal Eye Images

Abstract

Heart disease remains one of the most prevalent causes of death worldwide, with heart attacks accounting for a large proportion of morbidity and mortality. Early detection and accurate risk prediction are essential for prevention and timely medical intervention. Traditional diagnostic approaches often depend on invasive procedures or detailed clinical records, which may not always be practical or available. This research presents a novel, non-invasive approach for heart attack risk prediction using retinal fundus images. The retinal vasculature serves as a strong indicator of systemic cardiovascular health, and features such as vessel morphology, arteriovenous ratios, and tortuosity patterns are extracted for analysis. Conventional systems typically apply machine learning techniques such as AdaBoost and Convolutional Neural Networks (CNNs), relying on tabular medical datasets that include parameters like age, cholesterol, and blood pressure. Although effective, these methods do not fully exploit the rich visual biomarkers present in retinal images. The proposed system employs a deep learning model based on Recurrent Neural Networks (RNNs). High-resolution retinal images are transformed into sequential feature vectors representing vascular patterns across spatial regions. The RNN is designed to capture both spatial and temporal dependencies within these features, leading to more robust predictions of heart attack risk. Experimental results on publicly available retinal image datasets demonstrate that the RNN-based framework achieves an overall accuracy of 98%. These results outperform traditional machine learning classifiers, establishing the superiority of deep learning-based retinal analysis for cardiovascular risk prediction.

Introduction

Cardiovascular diseases, especially heart attacks, are a leading global cause of death, making early risk detection essential for prevention and treatment. Traditional diagnostic methods are often invasive, costly, or inaccessible in remote areas. Retinal imaging offers a promising non-invasive alternative since the retina’s blood vessels reflect systemic cardiovascular health. Changes in retinal vessel features—such as caliber, arteriovenous ratio, and tortuosity—correlate with heart disease risk.

This research proposes a novel heart attack risk prediction system using retinal fundus images analyzed by Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory (LSTM) models. Unlike CNNs, RNNs can model sequential and temporal dependencies, allowing them to capture complex spatial relationships and progression patterns in retinal vascular features by converting spatial data into sequential inputs.

The system preprocesses retinal images by enhancing contrast, reducing noise, and segmenting blood vessels to extract relevant features. These features are organized into sequences representing different retinal zones or vessel paths, which the LSTM model analyzes to predict heart attack risk levels (low, medium, high). The model can also integrate clinical data such as age or blood pressure to improve accuracy. A user-friendly interface supports real-time image input and risk visualization.

Extensive literature shows similar advances where deep learning models, especially hybrid CNN-RNN architectures and attention mechanisms, have improved cardiovascular risk prediction from retinal images. The system is evaluated on annotated retinal datasets, such as the Indian Diabetic Retinopathy Image Dataset (IDRiD), using metrics like accuracy, precision, recall, and AUC. It offers a cost-effective, accessible screening tool ideal for primary care and resource-limited settings, bridging ophthalmology and cardiology through AI.

Conclusion

Heart disease and heart attacks remain among the most significant global health challenges, making early detection and risk prediction a clinical priority. This research explored a novel, non-invasive approach to heart attack risk prediction using retinal fundus images, leveraging the fact that retinal vasculature reflects systemic cardiovascular health. Unlike conventional models that depend on structured clinical datasets (age, cholesterol, blood pressure, etc.), the proposed method integrates medical imaging with deep learning to extract and analyze vascular biomarkers. By transforming retinal images into sequential feature representations and employing a Recurrent Neural Network (RNN), the system effectively captures both spatial and temporal dependencies in vascular patterns. Experimental evaluation demonstrated that the RNN-based model achieved an accuracy of 98%, outperforming traditional machine learning approaches such as AdaBoost and CNNs. These findings highlight the potential of retinal imaging combined with advanced deep learning techniques as a cost-effective, scalable, and non-invasive solution for early heart attack risk prediction. With further validation on larger and more diverse datasets, this approach could be translated into real-world screening tools, enabling preventive healthcare interventions and reducing the global burden of cardiovascular disease.

References

[1] Zhang, Y., Li, W., & Chen, H. (2024). Attention-based convolutional neural networks for identifying heart failure risk from fundus images. Journal of Biomedical Imaging, 45(2), 112–120. [2] Liu, M., Zhao, K., & Wu, J. (2024). Hybrid CNN-RNN framework combining retinal imaging and patient metadata for improved heart disease prediction. Medical Image Analysis, 38(1), 76–85. [3] Alvarez, L., Gomez, F., & Ortega, M. (2024). Using recurrent neural networks to track retinal feature progression for early ischemic heart disease warning. Artificial Intelligence in Medicine, 138, 102452. [4] Desai, S., Nair, R., & Bansal, P. (2024). Ensemble RNN-XGBoost models for coronary artery disease prediction from retinal arteriolar narrowing. Pattern Recognition in Medical Imaging, 48(4), 91–99. [5] Singh, A., Verma, R., & Mehta, P. (2024). Interpretability of heart risk prediction via bidirectional RNNs and saliency maps on vascular trees. Neural Networks in Healthcare, 22(3), 178–189. [6] Yamamoto, N., Tanaka, H., & Ito, S. (2024). Augmenting RNN training for rare heart conditions with GAN-generated synthetic retinal data. Medical Image Synthesis, 6(1), 58–66. [7] Roy, S., Banerjee, K., & Saha, R. (2024). Sequence-to-sequence RNNs for longitudinal heart risk monitoring using retinal data. Journal of Digital Health Analytics, 4(3), 133–141.* [8] Wang, S., & Zhao, Y. (2025). Transformer-enhanced RNNs for myocardial infarction risk prediction from retinal vessel sequences. IEEE Transactions on Medical Imaging, 44(3), 512–523. [9] Patel, R., Sharma, D., & Gupta, A. (2025). Sequential LSTM modeling on segmented retinal vessels for heart attack risk assessment. Computers in Biology and Medicine, 159, 106721. [10] Kim, J., Park, E., & Lee, D. (2025). Real-time dynamic risk prediction using GRUs and retinal scan biomarkers. Scientific Reports, 15(1), 1445. [11] Chen, Y., & Huang, Z. (2025). Multi-modal fusion of retinal image sequences and electronic health records for improved heart risk prediction. Journal of Medical Data Science, 12(2), 88–97.* [12] Murthy, V., Rajan, S., & Kumar, T. (2025). Recurrent autoencoders for correlating retinal hemorrhages with cardiovascular risk. Computational Cardiology, 19(1), 67–75. [13] Islam, T., Chowdhury, M., & Rahman, H. (2025). Mobile-based lightweight RNN for heart risk classification from retinal images. Mobile Health Technologies, 10(2), 201–210 [14] Thomas, R., Joseph, A., & Iyer, S. (2025). Deep learning-based temporal vessel analysis surpasses traditional heart risk scoring. International Journal of Cardiovascular AI, 9(2), 121–130. [15] R.K. Bansal – Strength of Materials.P. Khurmi – Machine DesignASME Boiler and Pressure Vessel Code (for pressure piping and vessels). [16] Bullitt, E., Zeng, D., Gerig, G., Aylward, S., Joshi, S., Smith, J. K., ... & Lin, W. (2003). Vessel tortuosity and brain tumor malignancy: a blinded study. Academic Radiology, 10(12), 1378–1386. [17] The formula is from Long Short-Term Memory (LSTM) networks [18] The equation is from the LSTM (Long Short-Term Memory) architecture, specifically the input gate equation [19] This is the cell state update equation in a Long Short-Term Memory (LSTM) network, a special type of Recurrent Neural Network (RNN). [20] Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 [21] Deep Learning Book (Goodfellow, Bengio, Courville, 2016 [22] Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735–1780 [23] Hosmer, D. W., Lemeshow, S., & Sturdivant, R. X. (2013). Applied Logistic Regression. John Wiley & Sons.Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. [24] https://developers.google.com/machinelearning/crash-course/classification/accuracy-precision-recall

Copyright

Copyright © 2025 A. Syeda Tameema Anhar, Dr. Dilshad Begum. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.