gms | German Medical Science

Purpose: Cardiovascular disease (CVD) is ubiquitous and imposes high costs and mortality for patients and healthcare systems globally. Early and tailored diagnosis and treatment modalities may reduce CVD deaths and spending.

We aim to improve early CVD detection by using deep learning to find and refine biomarkers specific to the early stages of different CVD subtypes, especially subtle or subclinical signs that appear in early CVD.

Methods: We performed a targeted literature review to find the latest cardio-oculomics deep learning algorithm (DLAs) that discover CVD biomarkers. RetiAGE is a novel DLA that uses retinal photographs to estimate biological age and risk level for CVD mortality.

We then examined 57,297 retinal fundal photographs (RFPs) from the UK Biobank (UKBB) after removing low-quality RFPs, duplicates, and CVD cases at baseline. We generated RetiAge scores of these samples and confirmed that the RetiAGE scores were statistically significant in forecasting CVD events by doing a multivariate Cox proportional hazards (CoxPH) regression analysis adjusted for age and sex. We then did subgroup analysis via univariate CoxPH analysis on subtypes and dividing the cohort into quartiles based on RetiAge scores.

Results: RetiAGE predicts specific CVD subtypes, especially valvulopathies, embolic stroke, and arrythmias. RetiAge can capture subclinical changes in the retina in early-stage CVD, especially in essential (primary) hypertension. Subgroup analyses showed that the trend of hazard ratios of CVD events increasing from quartiles 2 to 4 was generally present in most subtypes.

Conclusion: DLAs can detect subclinical changes to predict and stratify CVD morbidity. As CVD is heterogenous, adjusting the RetiAge score to specific subtypes may allow personalised CVD subtype risk ranking, which may improve health outcomes.