A Comprehensive Detection of Cardiovascular Risks Utilizing Deep Learning & Ensemble Learning Methods Through Retinal Fundus Images

Cardiovascular diseases (CVDs) have evolved as a critical public health challenge driving high mortality rates and causing fatal physical damages. The CVDs have immense adverse impacts on public health as well as on the economy. Therefore, addressing and mitigating the influences caused by CVDs requires careful attention in the effective detection of CVDs. Thankfully, Deep Learning (DL) and Machine Learning (ML) approaches exhibit significant progress in detecting CVDs. In this paper, DL and ML approaches are implemented to detect the risk of 3 CVDs including Stroke, Congestive Heart Failure (CHF), and Hypertension, from retinal fundus images (RFIs). DL approaches, specifically, Convolutional Neural Network (CNN)-based Transfer Learning (TL) models including VGG16, VGG19, Xception, DenseNet121, DenseNet169, and DenseNet201 are considered feature extractors (FEs). A custom DL-based model has been utilized as a classifier. Additionally, a set of ML classifiers such as Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) has been employed. Moreover, an Ensemble Learning (EL) approach has been implemented. RFIs have been accumulated from 8 datasets. The EL approach emerged as the best performer in detecting both the risk of Stroke and CHF (accuracy: 94.97%, recall: 94.69%, precision: 95.64%, f1-score: 95.19%, and AUC: 99.08%) as well as Hypertension (accuracy: 96.02%, recall: 93.73%, precision: 93.78%, f1-score: 93.76%, and AUC: 98.09%). The Local Interpretable Model-agnostic Explanations (LIME) has been implemented as an Explainable Artificial Intelligence (XAI) technique to yield feature visualization. This endeavor can be considered an outstanding contribution to the fruitful detection of CVDs.