Automated retinal microvascular quantification as a predictor of cardiovascular disease risk in later life

Retinal vessels (both arteries and veins) on the back of the eye are easily imaged using fundus cameras. The shape and size of retinal vessels (particularly arteries), have been related to risk of cardiovascular disease in later life and may provide valuable prediction of individuals at high risk of disease. Recent studies have also shown that retinal vessel size and shape are associated with cardiovascular risk factors (including blood pressure and blood cholesterol) from an early age, suggesting that these measures may be non-invasive markers of the health of blood vessels.

Advances in imaging and computational methods are providing opportunities to image and analyse retinal vessels with increasing precision. However, to date these methods are not fully automated (with respect to the relevant datasets) and rely heavily on operator involvement, especially to distinguish arteries from veins (which may change in size and shape differently in response to disease). These problems limit the application of retinal imaging in large population studies of cardiovascular disease and other key chronic diseases in later life, particularly ocular disease (macular degeneration and glaucoma) and type 2 diabetes. This proposal aims to fully automate identification and characterisation of vessels in retinal images, including identification of arteries and veins, in large studies of middle aged adults (including UK Biobank - UKBB). Images (nearly 150,000 images from over 70,000 participants) will be used to refine our existing approaches to measurement of retinal vessel size and shape (which include both measures of vessel width and tortuosity, i.e., a measure of twisting / turning), and to examine their association with risk of cardiovascular disease in later life. Middle age adults are often seen by family doctors and given a risk score for cardiovascular disease (i.e., risk of heart attack), based on a number of simple measures (e.g., age, gender, blood pressure and blood cholesterol); this often determines whether treatments (e.g., BP medications, statins) are prescribed.

The project brings together a well-established, multidisciplinary group with the shared aim of developing reliable, automated and efficient retinal image analysis software, which is freely available, generating a rich characterization of the retinal vessels in large numbers retinal images. Innovative use of the large-scale data generated from thousands of participants, will be ensured by the inclusion of appropriate statistical expertise within the group. Current image analysis is underway using software developed within the group, and this proposal seeks to develop the software into a usable form available to all. The potential scientific value of the project is appreciable: (1) Screening: retinal vessel characteristics may provide measures of abnormality that are predictors of overall vascular health and serious vascular disease in later life. In addition to established screening tools, this may identify "at-risk groups" in need of earlier treatment and therefore reduce the complications associated with cardiovascular events. (2) New resources: the validated retinal measurements generated will become themselves data to add to these cohorts, enabling further investigation by the wider scientific community. (3) Collaboration: Combination of specialist groups is expected to generate a step-change improvement in performance of these software tools. (4) Long-term scientific value of the work is assured by making the software developed by the group, available for public use. This ensures that the value of examining retinal vessels as a predictor of disease risk and outcome (both circulatory and ocular disease) is realised by the widest possible scientific community.

Overt changes in retinal vessels (particularly arterioles) have long been associated with disease. Growing evidence, including longitudinal evidence, suggests morphological changes in retinal vessels (including changes in tortuosity and width) are early physio-markers of cardio-metabolic risk and outcome (as well as other disease processes). Despite rapid advances in image capture and computational assessment of retinal vessels, considerable operator involvement limits investigation in large population studies. This project brings together a multidisciplinary group, with the shared aim of developing reliable, automated, efficient, freely available retinal image analysis (RIA) software, generating a rich quantitative characterization of the retinal vasculature in large volumes of fundus images. Previous work carried out by the investigators, including width and tortuosity measures, will ensure appropriate and novel use of the large-scale data generated from RIA from thousands of participants. An automated approach to retinal image analysis building on software already developed will be extensively validated against clinical assessment of vessel features, and applied to retinal images from large MRC funded prospective cohort studies in adulthood, including the European Prospective Investigation of Cancer (EPIC Norfolk - a richly phenotyped data source with 8623 participants), and UK Biobank (a prospective study of 500,000 middle aged adults, 70,000 with retinal images using the same image capture method as EPIC, i.e., a Topcon non-mydriatic fundus camera), allowing its performance in prediction of CVD events in middle aged and older age groups to be examined. Examination in the larger UK Biobank cohort will allow both cross-sectional and future prospective associations with CVD risk to be explored definitively (due to improved statistical power). Re-examination of the cohort (in 10,000 participants) will also allow analysis of repeated retinal imaging over time.

Initial work will be research led, establishing definitive associations between retinal vessel morphology and cardio-metabolic precursors and outcomes. Immediate availability of EPIC data will allow cross sectional associations to be examined within this existing richly phenotyped data set (especially as retinal imaging was only carried out in the latest phase of examination). Data from UK Biobank will allow associations observed in EPIC Norfolk to be replicated with a prior hypothesis, providing strong evidence of association. While associations with disease precursors (cardio-metabolic risk factors) may suggest causal associations, further follow-up of cohort participants will allow causality to be inferred if the temporal sequence of association is observed. Hence, this proposal will establish the role of retinal vessel assessment in risk prediction of cardio-metabolic events, and the potential contribution retinal vessel assessment has to current risk prediction scores. Retinal imaging is non-invasive, and cheap. Hence, assessment may well have a role to play in identifying those at high risk, to ensure prophylactic measures can be taken before the onset of overt disease. Such an approach may have potential savings to the NHS / health service in terms of delayed treatment costs, by extending disease free survival.