Heart Attack Risk Factor Analysis and Prediction Using Retinal Eye Images, Digital Image Processing, and Deep Learning Algorithms

Abstract

Myocardial infarction continues to rank as the world's single most lethal medical event, responsible for an estimated 17.9 million fatalities annually. Ironically, much of that burden is preventable, provided that susceptible individuals are identified before their first acute episode. The difficulty is that existing risk-scoring methods, however scientifically sound, demand fasting blood draws and lipid assays that large segments of the global population simply cannot obtain. Authors set out to design a screening tool that bypasses these logistical barriers entirely. The approach harnesses a medically underused fact: the inner surface of the eye, specifically its network of tiny blood vessels, undergoes measurable structural changes driven by the same pathological forces that damage the coronary arteries. A photograph of the retina, therefore, encodes cardiac risk information that can be read by a trained algorithm. Authors built a three-stage learning architecture for this task. In the first stage, an Inception-v3 convolutional network—initialized on a broad natural-image corpus and subsequently specialized on fundus photographs—extracts a compact numerical representation of each patient's retinal morphology. In the second stage, Fuzzy C-Means organizes patients into graded risk strata whose boundaries reflect the probabilistic, spectrum-like nature of cardiovascular disease rather than imposing arbitrary hard thresholds. In the third stage, a two-layer Long Short-Term Memory network interprets the combined representation and delivers a binary verdict: elevated or non-elevated cardiac risk. Fourteen structured physiological aisbiochemical variables from the UCI Cleveland Heart Disease collection are woven into the feature space alongside the image-derived descriptors, enriching the model with complementary clinical context. The trained system is wrapped in a Flask web application accessible by patients and clinicians through role-differentiated dashboards. Tested against a held-out partition of the Cleveland dataset, the architecture attained 91.3 % classification accuracy and an area under the receiver-operating curve of 0.94. These figures exceed those of logistic regression (78.4%, AUC 0.82), conventional deep networks (83.6%, AUC 0.88), and all prior published methods on this benchmark. A systematic component ablation confirms that retinal imaging supplies predictive information orthogonal to structured clinical data.