Beyond vision: leveraging retinal biomarkers to predict cardiovascular disease.
Lead Research Organisation:
The University of Manchester
Department Name: School of Medical Sciences
Abstract
One in three deaths are caused by cardiovascular disease (CVD), and three-quarters of early CVD could be prevented. To avoid untimely deaths, those at risk must be identified early to allow targeted preventative strategies. CVD risk is currently assessed using calculators based on health information and blood tests. The hypothesis driving this proposal is that the accuracy of these assessments can be enhanced using eye scans.
The eye is the only part of the body where the nerves and blood vessels can be directly visualised. The appearance of these structures provides important insights into an individual's cardiovascular health. Two types of scans can be used on the eye: (1) photographs of the inner lining of the eye (the retina), known as fundus photographs, and (2) more detailed 3D scans of the eye known as 'optical coherence tomography' (OCT). Using these, almost every high street optometrist can now take low-cost, risk-free, sophisticated eye scans in just a few seconds.
At present, eye scans do not form part of CVD risk assessment in primary care. The tool currently used is a questionnaire-style risk calculator which considers health and lifestyle information called "QRISK3". An emerging tool proposed as a future risk assessment solution is "polygenic scores". Polygenic scores are measures of disease risk based on genetic information, and these are currently being explored in research settings. Although it is known that eye scans can be used to predict CVD, it is not known how much more information they offer compared to these established and emerging tools. Furthermore, it is not clear why certain features in eye scans aid prediction of CVD, or if it will be possible to accurately predict CVD in patients who have eye disease. If these problems can be addressed, it is possible that eye scans could be used in the community to identify people in need of medical attention to prevent CVD.
To address these problems, I have outlined three key objectives:
1. To explore why the presence of certain OCT/fundus photograph features predict CVD risk.
2. To measure the level of improvement to the accuracy of CVD predictions obtained by adding eye scan data to an established (QRISK3) and an emerging (polygenic score) risk tool.
3. To evaluate how well the developed tools function in people with common eye diseases and begin to address any effects on prediction accuracy.
To perform my research, I will use health data from over 115,000 people available via biobanks and collaborators. I will begin by comprehensively studying what biological factors account for the role of known eye markers in CVD. Next, I will develop OCT plus fundus photograph-based CVD prediction tools using artificial intelligence. I will compare the accuracy of CVD predictions derived from QRISK3, polygenic scores, and an enhanced tool that incorporates OCT and fundus photographs. It is known that age-related macular degeneration (AMD) reduces the accuracy of risk assessments based on eye scans, so I will train my software to identify AMD from the images and mitigate this effect. Finally, I will test the accuracy of my software in two other common eye diseases, cataract and diabetic retinopathy, to assess how well my tool works in people with these problems.
This work will represent progress toward cheap, accessible, accurate, opportunistic screening to identify those at risk of future CVD.
The eye is the only part of the body where the nerves and blood vessels can be directly visualised. The appearance of these structures provides important insights into an individual's cardiovascular health. Two types of scans can be used on the eye: (1) photographs of the inner lining of the eye (the retina), known as fundus photographs, and (2) more detailed 3D scans of the eye known as 'optical coherence tomography' (OCT). Using these, almost every high street optometrist can now take low-cost, risk-free, sophisticated eye scans in just a few seconds.
At present, eye scans do not form part of CVD risk assessment in primary care. The tool currently used is a questionnaire-style risk calculator which considers health and lifestyle information called "QRISK3". An emerging tool proposed as a future risk assessment solution is "polygenic scores". Polygenic scores are measures of disease risk based on genetic information, and these are currently being explored in research settings. Although it is known that eye scans can be used to predict CVD, it is not known how much more information they offer compared to these established and emerging tools. Furthermore, it is not clear why certain features in eye scans aid prediction of CVD, or if it will be possible to accurately predict CVD in patients who have eye disease. If these problems can be addressed, it is possible that eye scans could be used in the community to identify people in need of medical attention to prevent CVD.
To address these problems, I have outlined three key objectives:
1. To explore why the presence of certain OCT/fundus photograph features predict CVD risk.
2. To measure the level of improvement to the accuracy of CVD predictions obtained by adding eye scan data to an established (QRISK3) and an emerging (polygenic score) risk tool.
3. To evaluate how well the developed tools function in people with common eye diseases and begin to address any effects on prediction accuracy.
To perform my research, I will use health data from over 115,000 people available via biobanks and collaborators. I will begin by comprehensively studying what biological factors account for the role of known eye markers in CVD. Next, I will develop OCT plus fundus photograph-based CVD prediction tools using artificial intelligence. I will compare the accuracy of CVD predictions derived from QRISK3, polygenic scores, and an enhanced tool that incorporates OCT and fundus photographs. It is known that age-related macular degeneration (AMD) reduces the accuracy of risk assessments based on eye scans, so I will train my software to identify AMD from the images and mitigate this effect. Finally, I will test the accuracy of my software in two other common eye diseases, cataract and diabetic retinopathy, to assess how well my tool works in people with these problems.
This work will represent progress toward cheap, accessible, accurate, opportunistic screening to identify those at risk of future CVD.
Technical Summary
Cardiovascular disease (CVD) causes a third of deaths globally. Most of these deaths are due to stroke or myocardial infarction (MI). Notably, it is estimated that 75% of premature CVD is preventable. Although dramatic declines in CVD mortality were achieved in the late 20th century, this decline has not only plateaued, but in England CVD deaths are at a 14-year high. It is estimated that the 10-year cumulative incremental net monetary cost associated with a persistent plateauing of CVD will be approximately £54 billion in England and Wales alone. Meanwhile, our prevention strategies are faltering. Although the NHS health check aims to assess and mitigate CVD risk in those aged 40-74, attendance is falling year-on-year, and in 2022/23 stood at just 38.9% in England. It is therefore important that we seek new ways to identify those at risk of CVD. Retinal imaging-based CVD risk assessment represents a promising avenue to begin to address this challenge.
It takes seconds to acquire a retinal image, with little cost, and no risk to the patient. Vasculopathy and metabolic disease manifest in the microvasculature of the retina at an early disease stage and provide sensitive indicators of future vascular event. To date, we have been unable to leverage this knowledge to benefit the population due to limitations of technology and accessibility. However, advances in imaging technology have now enabled widespread cost-effective evaluation of retinal parameters in both hospital and optometry settings.
During this fellowship, I will study the role of retinal imaging in predicting common cardiovascular disorders, including transient ischaemic attack (TIA), stroke, and myocardial infarction. Using data from the UK Biobank and a set of independent cohorts, I will explore how features extracted from fundus photographs and ocular coherence tomography images can enhance cardiovascular disease risk assessments. I will develop predictive models using deep learning techniques, integrating a combination retinal imaging, genetic information, and clinical data. I will validate my models in independent cohorts, and specifically test model performance in those with ocular diseases. I expect age-related macular degeneration to affect model performance, and so I will develop a proof of principle mitigation strategy to improve model performance in those with this disease. Finally, I will explore the biology linking retinal features to CVD risk using genetics-based techniques including Mendelian randomisation.
In summary, during this project I will develop biologically explainable retinal imaging-based models for the prediction of cardiovascular disease. I will externally validate my models, including in those with ocular disease. The successful completion of this project has the potential to influence primary and secondary prevention of cardiovascular disorders.
It takes seconds to acquire a retinal image, with little cost, and no risk to the patient. Vasculopathy and metabolic disease manifest in the microvasculature of the retina at an early disease stage and provide sensitive indicators of future vascular event. To date, we have been unable to leverage this knowledge to benefit the population due to limitations of technology and accessibility. However, advances in imaging technology have now enabled widespread cost-effective evaluation of retinal parameters in both hospital and optometry settings.
During this fellowship, I will study the role of retinal imaging in predicting common cardiovascular disorders, including transient ischaemic attack (TIA), stroke, and myocardial infarction. Using data from the UK Biobank and a set of independent cohorts, I will explore how features extracted from fundus photographs and ocular coherence tomography images can enhance cardiovascular disease risk assessments. I will develop predictive models using deep learning techniques, integrating a combination retinal imaging, genetic information, and clinical data. I will validate my models in independent cohorts, and specifically test model performance in those with ocular diseases. I expect age-related macular degeneration to affect model performance, and so I will develop a proof of principle mitigation strategy to improve model performance in those with this disease. Finally, I will explore the biology linking retinal features to CVD risk using genetics-based techniques including Mendelian randomisation.
In summary, during this project I will develop biologically explainable retinal imaging-based models for the prediction of cardiovascular disease. I will externally validate my models, including in those with ocular disease. The successful completion of this project has the potential to influence primary and secondary prevention of cardiovascular disorders.
