Access to cardiac MR imaging centres is limited in developing countries, scanners being expensive and difficult to acquire. Even in developed nations like the UK, there is a limited number of cardiac MR imaging centres and waiting lists are long.
A multidisciplinary study, led by Dr Nishant Ravikumar, Lecturer in Computer Science, and Prof. Alejandro Frangi who holds the Diamond Jubilee Chair in Computational Medicine, at the University of Leeds, looked at whether an alternative, more cost-effective, and scalable solution could be found to detect indicators of cardiac disease, especially of Myocardial Infarction (MI).
Doctors have, for a while now, recognised that changes to the tiny blood vessels in the retina are indicators of broader vascular disease, including problems with the heart. Research conducted by Google Help in 2018-19 demonstrated it is possible to predict cardiac risk factors using Fundus (retinal) photography. AI for retinal screening thus has the potential to revolutionise the way patients are regularly screened for signs of heart disease.
“We wanted to take it one step further than the Google study, to see if we could, instead of just predicting risk factors, predict MI events using Fundus photography images and basic patient information,” explains Dr. Nishant Ravikumar, who has a background in Biomedical Engineering and Machine Learning. The first step was to effectively combine information from both cardiac imaging and retinal imaging, by learning a common representation from the two sets of data (retinal images and cardiac images). That way the representation learned by the model could be used to directly predict certain characteristics of the heart or adverse future events such as myocardial infarction, given retinal imaging alone.”
The team used the UK Biobank database, a biomedical database and research resource, that has in-depth genetic and health information from half a million healthy volunteers. The resource had information on 29,000 people who had had a cardiac MRI scan and 74,000 individuals with OCT/Fundus imaging. Dr Ravikumar and his team extracted this data and found a crossover of 11,000 healthy people which had both MRI and retinal imaging. They then examined how many participants did not have any diagnosed CV disease at the time of their imaging. As the imaging was done in 2014, several of the volunteers had since had MI events, which the research was attempting to predict. Pre-processing steps were taken to remove cofounding variables, including diabetes and other pre-existing conditions – in order to prevent bias in the findings.
Biobank data is acquired from only a small number of scanning centres and therefore lacks real world variability in the Fundus images, so reviewers asked the LIDA team to undertake a replication study using data from another cohort. “That’s where things got pretty interesting,” says Dr Ravikumar. “Using another dataset from the US National Institute of Health, which had retinal images along with follow-up patient information, we found we can predict MI using retinal images in the US and the UK, to a reasonable degree of accuracy, up to three years into the future from when the images were acquired.”
Envisaged as a proof-of-concept study, this ground-breaking work demonstrates the possibility of screening for, and identifying, patients at risk of future MI events based on just retinal imaging. Additional work and rigorous testing will potentially culminate in a real-world clinical trial. And because the model needs only Fundus photography, which is routinely obtained by opticians during check-ups and at eye clinics, this could be a very low-cost way to screen for future heart attacks. The research also included a series of experiments that showed only a modest number of additional clinical measurements are needed to accompany the images, basic demographic variables such as age, gender, and blood pressure to predict the risk of developing MI.
Identifiable cardiac diseases, complications or end points such as MI or cardiomyopathy, all result from cascades of events happening over a period of years. What is generally known as Cardiometabolic Syndrome (CMS) involves a build-up of fat in the liver and pancreas, along with the development of diabetes. Interplay between diabetes and atherosclerosis means a person’s vascular health is then made much worse, ultimately leading to a heart attack or other vascular or CV complications.
“Where we see our work benefiting society is in screening for cardiac disease at a low cost, meaning more people at risk can be identified early on, with measures then taken to reduce their risk of suffering further complications a few years down the line,” explains Dr Ravikumar. “This study highlights the fact that there is this systemic response happening in the body over time, meaning you can find indicators of CV health from a simple, two-dimensional photograph of the eye.”
The final manuscript detailing this pioneering work has been accepted for publication in Nature Machine Intelligence and is due to appear in early 2022.
For further information: