EyeQ: Democratizing Cardiovascular Risk Screening with a Reproducible, Explainable, and Low-Cost Mobile-Based Approach

Overview

From time-consuming diagnostics to inaccessible appointments, understanding personal cardiovascular health is a challenging process that has become a practical tedium for many. Because limited access to advanced cardiovascular screening persists, many individuals rely on unreliable online risk calculators; beyond false negatives, these tools leave asymptomatic carriers of disease, poised at the threshold of cardiac crisis, wholly unprotected. To address this gap, our team constructed a scalable machine learning model capable of predicting cardiovascular risk using only retinal fundus images and designed a complementary camera system. Our system provides several benefits: (1) the novel use of semi-supervised learning ensures the model remains reproducible for researchers with limited access to large labeled datasets while maintaining accuracy; (2) the incorporation of explainable AI (Grad-CAM) enhances transparency in clinician-patient interactions; and (3) the development of a low-cost smartphone-based retinal fundus camera expands accessibility. An unsupervised model based on SimCLR architecture was first trained on the Rotterdam AIROGS dataset. Learned representations were stored and transferred to a supervised module trained on the China-CIMT dataset to predict carotid intima-media thickness (CIMT). To ensure reproducibility, both datasets utilized were publicly available. Predicted CIMT values were categorized into risk groups (>0.9 mm high risk; ≤0.9 mm low risk). Our model demonstrated stable convergence during unsupervised training (~2.38 validation loss) and successfully learned retinal features predictive of CIMT, achieving ~0.129 MAE and ~0.8138 AUROC. Overall, this work highlights the potential of accessible retinal imaging and semi-supervised learning to support scalable, early cardiovascular risk screening.