Eyescan opens the window to heart health

Scientists at the University of Leeds have successfully trained an artificial intelligence (AI) system to detect signs of heart disease from retinal scans routinely performed by opticians and ophthalmologists at a relatively low cost. ..

“Wouldn’t it be great if we could use our eyes as a window to assess cardiovascular health?” Asked Dr. Alejandro Frangi, senior author of the study and professor of computational medicine at the University of Leeds. rice field.

In a study published in Nature Machine Intelligence title, “Prediction of myocardial infarction with retinal scan and minimal personal information” Frangi and his team have shown that the AI ​​system can analyze retinal images to predict whether a patient is at risk for a heart attack next year with 70-80% accuracy.

This highly translatable system, which can be introduced into the clinic to predict the risk of myocardial infarction, provides basic biological insights that separate the relationship between retinal microvessels and the anatomical abnormalities of coronary heart disease. Inspired by recent attempts to extract using a deep learning model, information on cardiovascular risk factors from features of the retinal fundus such as optic nerve papilla and vascular structure. Previous studies have shown that deep learning can predict smoking status, blood pressure, age, and other cardiovascular risk factors from retinal scans.

“The potential to develop an early risk score for a heart attack can be used to refer patients to a cardiologist through an ophthalmology clinic, optician, or potentially self-management,” Frangi said. Similar approaches have been developed to detect glaucoma via smartphone-based deep learning systems. Deep learning is a set of complex algorithms that allow a computer trained to identify patterns of data to make predictions.

The current study uses AI prediction models trained in more than 70,000 individuals in the UK Biobank Cohort and tested in an independent US cohort (AREDS cohort) consisting of UK Biobank data and 3,000 patients. increase.

“Our proposal differs in several respects compared to previous work by DeepMind,” says Frangi.

Early studies by DeepMind focused on diagnosing eye disease rather than heart disease from retinal scans. In addition, previous studies by Google Research predicted cardiovascular risk factors such as smoking status, blood pressure, and age, but did not predict heart events such as heart attacks from retinal images.

“Our method allows us to use both fundus and heart images during the training phase, but risk estimation requires only fundus images,” says Frangi. Estimating cardiac parameters such as ejection fraction in combination with retinal anatomy enhances the explanatory power of our unique new approach.

In this study, researchers used retinal images and patient demographic data to estimate the mass of the left ventricle of the heart and used the end-diastolic volume of the left ventricle to predict myocardial infarction.

“The underlying technology we use is called a multi-channel variational autoencoder,” says Frangi.

Researchers have trained multi-channel variational autoencoders and deep regressor models to estimate these cardiac attributes from retinal images and demographic data.

In this paper, the authors presented two alternative models for predicting cardiac events with relatively comparable performance. “We assume that additional information about the age and gender of the individual under test is minimal,” Frangi said. “The second is relatively low cost, but incorporates additional biomarkers that are primarily available in specialized ophthalmology clinics.”

“Our results show that retinal imaging available at all opticians and ophthalmology clinics can identify patients at high risk for future myocardial infarction,” the authors say.

Dr. Sven Plein, a professor of cardiovascular imaging at the University of Leeds and one of the authors of the paper, said: “AI systems are a great tool for unraveling complex patterns that exist in nature — complex patterns of retinal changes related to changes in the heart.”

The team has adopted several research directions to follow up on current work. “We would like to develop a prospective clinical trial to demonstrate the predictive value of this technology in the general population with diverse demographic profiles under NHS operational conditions compared to UK Biobank. “Frangi said. “We also want to understand comorbidities that affect performance as confounders.”

Currently, diagnostic signatures of the heart, such as the size of the patient’s left ventricle and pump efficiency, can only be determined by ultrasonography or magnetic resonance imaging. These diagnostic tests can be expensive and require a visit to the hospital. The use of deep learning approaches to analyze retinal scans has the potential to revolutionize medical care through early and inexpensive detection of signs of heart disease.