Scientists at the University of Leeds have successfully trained artificial intelligence (AI) systems to detect signs of heart disease from retinal scans that are comparatively cheap and routinely performed by opticians and eye clinics.
“Wouldn’t it be great if we could use our eyes as a window to assess our cardiovascular health?” asked Alejandro Frangi, PhD, senior author of the study and professor of computational medicine at the University of Leeds.
In a study published in Nature Machine Intelligence titled, “Predicting myocardial infarction through retinal scans and minimal personal information,” Frangi and his team showed that the AI system can analyze retinal images to predict if patients are at risk of a heart attack over the next year, with an accuracy rate of 70 to 80 percent.
This highly translatable system that can be deployed in clinics to predict the risk for myocardial infarction was inspired by basic biological insights that teased apart the relationship between anatomical abnormalities in retinal microvascular and coronary heart disease, and more recent attempts that used deep learning models to extract information on cardiovascular risk factors from features of the retinal fundus such as the optic disc and vascular architecture. Earlier work had shown deep learning can predict smoking status, blood pressure, age, and other cardiovascular risk factors from retinal scans.
“The possibility of developing an early risk score of heart attack could be used to refer patients to a cardiologist through eye clinics, opticians, or potentially self-management,” said Frangi. Similar approaches have been developed for the detection of glaucoma through a smartphone-based visual field deep learning system. Deep learning is a set of complex algorithms that enable computers to make predictions once trained to identify patterns in data.
The current study uses an AI predictive model that was trained on over 70,000 individuals from the U.K. Biobank Cohort and tested using U.K. Biobank data and on an independent U.S. cohort of 3,000 patients (AREDS cohort).
“Compared to previous work by DeepMind, our proposal differs in several ways,” said Frangi.
Earlier work by DeepMind focused on diagnosing eye diseases and not cardiac disease from retinal scans. In addition, earlier work by Google Research predicted cardiovascular risk factors such as smoking status, blood pressure, age, and not cardiac events such as heart attacks from retinal images.
“Our method can use imaging of both the fundus and the heart during the training phase yet only require fundus images during risk inference,” said Frangi. Estimating cardiac parameters such as ejection fraction in conjunction with retinal anatomical features enhances the explanatory power of the unique, new approach.
In the work, the investigators used retinal images and the patients’ demographic data to estimate the mass of the left ventricular chamber of the heart, and the left ventricular end-diastolic volumes to predict myocardial infarction.
“The underpinning technology we use is called multi-channel variational autoencoders,” said Frangi.
The researchers trained a multi-channel variational autoencoder and a deep regressor model to estimate these cardiac attributes from retinal images and demographic data.
In the paper, the authors presented two alternative models for predicting cardiac events with relatively comparable performance. “One assumes minimal additional information in the age and sex of the individual being tested,” said Frangi. “The second incorporates additional biomarkers that are relatively low-cost but mostly available in specialized eye clinics.”
“Our results indicate that one could identify patients at high risk of future myocardial infarction from retinal imaging available in every optician and eye clinic,” the authors noted.
Sven Plein, PhD, professor of cardiovascular imaging at the University of Leeds and one of the authors of the paper said, “The AI system is an excellent tool for unraveling the complex patterns that exist in nature, and that is what we have found here—the intricate pattern of changes in the retina linked to changes in the heart.”
The team is adopting several directions of investigation to follow up on the current work. “We’d like to develop a prospective clinical trial to demonstrate the predictive value of this technology in the general population with a diverse demographic profile compared to UK Biobank and under NHS operational conditions,” said Frangi. “We would also like to understand the comorbidities affecting performance as confounders.”
At present diagnostic signatures of the heart such as the size and pumping efficiency of a patient’s left ventricle can only be determined through echocardiography or magnetic resonance imaging. These diagnostic tests can be expensive and require hospital visits. The use of deep learning approaches to analyze retinal scans could revolutionize healthcare through early and inexpensive detection of signs of heart disease.
