Predicting major adverse cardiovascular events with colour fundus photograph in the AlzEye Study

Abstract

Purpose : Retinal images contain signs of systemic health conditions, such as cardiovascular diseases. Colour fundus photography is a promising risk stratification tool due to its affordability, wide-spread availability, and non-invasive acquisition. To investigate the utility of colour fundus photographs in analysing major adverse cardiovascular events (MACE), we developed an AI model based on our in-house foundation model to predict the three-year incidence of MACE.

Methods : The study cohort is from the Moorfields AlzEye project, a retrospective cohort study linking ophthalmic data of 353,157 patients, who attended Moorfields Eye Hospital between 2008 and 2018, with systemic health data from hospital admissions across England. 5,382 patients had retinal imaging within three years before developing MACE (ICD codes: stroke I63-I64, myocardial infarction I21-I22, heart failure I50, and atrial fibrillation I48). For each patient, we included the left retinal image from a single visit to avoid potential bias by inconsistent individual visits. The colour fundus photographs were preprocessed with AutoMorph (https://rmaphoh.github.io/projects/automorph.html). We split the patients into train, validation, and test sets with a ratio of 55%:15%:30%. The model was fine-tuned on training data and then evaluated on the test set. 95% confidence intervals (CI) were calculated by bootstrap.

Results : The developed model achieved an AUROC of 0.783 (95% CI 0.783 0.784) and AUPR of 0.754 (95% CI 0.753, 0.755). The F1-score of MACE prediction was 0.734 (95% CI 0.733, 0.734) and the sensitivity achieved 0.785 (95% CI 0.785, 0.786). The performance was significantly better than another competitive baseline we have developed using the architecture of transformers, with an AUROC of 0.723 (95% CI 0.721, 0.724), AUPR 0.692 (95% CI 0.69, 0.694), F1-score 0.665 (95% CI 0.663, 0.666), and sensitivity 0.674 (95% CI 0.672, 0.675).

Conclusions : The developed model performs well in predicting the incidence of MACE in three years, markedly outperforming the powerful compared method. This verifies that retinal images indeed include markers for complex health conditions and demonstrates the potential of identifying individuals with high risks with our model with colour fundus photographs.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.