Early research suggests promising use of artificial intelligence to predict the 10-year risk of death from a heart attack or stroke from a single chest X-ray.
The preliminary results were presented Tuesday at the annual meeting of the Radiological Society of North America. The research is in the final stages of drafting and has not yet been submitted for publication in any medical journal.
The researchers used nearly 150,000 chest X-rays to train an artificial intelligence program to identify patterns in the images that are linked to the risk of serious cardiovascular disease. They tested the program on a separate group of about 11,000 people and found a “significant association” between the level of risk predicted by the AI and the actual occurrence of major cardiovascular disease.
The clinical standard for cardiovascular disease risk analysis is the Atherosclerotic Cardiovascular Disease (ASCVD) Risk Score, a calculator that weights various patient data points that are highly associated with adverse cardiovascular events , including age, blood pressure and history of smoking.
Statin drugs are recommended for people with a 10-year risk of 7.5% or greater. The AI model uses the same risk thresholds as the established risk calculator, and early evidence suggests it works just as well.
“We have long recognized that X-rays capture information beyond traditional diagnostic findings, but we did not use this data because we did not have robust, reliable methods,” said Dr. Jakob Weiss, the principal investigator and associate radiologist with Massachusetts General Hospital and the AI in Medicine program at Brigham and Women’s Hospital at Harvard Medical School.
Sometimes the AI results match a traditional radiology report, but other times they pick up on things that may have been missed, he said.
“Some of these are anatomical changes that we would also be able to see with the naked eye and that make physiological sense. Let’s say there is high blood pressure or heart failure – these are findings that we can also see in a normal chest X-ray. But I think a lot of the information captured or extracted is embedded somewhere in the scan, but we as traditionally trained radiologists can’t understand it right now,” Weiss said.
“It has this black-box nature,” he said, which can sometimes make it difficult to communicate risks to patients without a clear explanation.
dr Donald Lloyd-Jones, Chair of Preventive Medicine at Northwestern University’s Feinberg School of Medicine and past president of the American Heart Association, was co-chair of the risk assessment panel when the ASCVD risk calculator was created in 2013 and a key figure in 2018 when the guidelines were updated were used to highlight the relationship between risk score and personal medical history.
He wasn’t involved in the new AI research but says it’s important to move the field forward.
“This is exactly the type of application that artificial intelligence is best suited for,” he said. “So we have to keep doing things like this to really understand if we can find patients in particular who would otherwise slip through the cracks. I think that’s where it could be most useful.”
But collecting all patient data points that go into the mainstream risk calculator is still critical – because they are actionable. And whether risk is calculated using a statistical formula or an AI model, the most successful outcomes still require personalized patient assessments.
“We don’t cure smoking with a chest X-ray. We really need to work with the patient to find ways to get them to quit,” Lloyd-Jones said. “The risk calculator is part of the risk assessment, but not the only part. It’s a process that engages both the patient and the doctor in a discussion about what the patient’s risk is and how much we think a statin would help them.”
For their research, Weiss and co-authors trained the AI using chest X-rays from participants in the National Cancer Institute’s prostate, lung, colon, and ovarian cancer screening study. It was tested on people who had undergone routine ambulatory chest x-rays at Mass General Brigham and were potentially eligible for statin therapy, with a median age of 60 years.
Additional research, including a controlled randomized trial, is needed to validate the deep learning model.