This is one of several deep learning algorithms developed by Mass General and Brigham and Women's to look at standard chest X-rays to use AI to perform a variety of opportunistic screenings of patients who get a chest X-ray for various reasons. They developed algorithms to look at chest X-rays for lung screenings, determining biological age of a patient to determine their longevity better than current methods, and now cardiac assessments. The AI is still in development, but represents a new trend in radiology AI that will likely see commercialization by several vendors over the next couple years.

Since chest X-rays are the most common radiological exam performed worldwide, AI development efforts are being concentrated on this exam type. The AI extract thousands of small radiomic data points that cannot be seen by humans.  

"We found that there is pretty robust information captured by the chest X-ray that we were not aware of and that we could not quantify, but with deep learning techniques we can now extract this information and use it for risk prognostication," Weiss said. 

The AI can predict the 10-year risk of death from a heart attack or stroke. The algorithm was trained using 150,000 chest X-rays from 40,643 patients, where the patient history and outcomes were know. The deep learning system was able to look at patient data and their X-rays and determine radiomic patterns in the images that were the same in patients who had similar outcomes. It was then tested on about 11,000 X-rays from other patients without any additional information and the AI was able to accurately determine the risks of these patients using the imaging data alone.

Weiss said that, in radiomics, it is not one or two simple things like seeing an enlarged heart that is evident to radiologist. It is a combination of thousands of small data points in the image that the AI can quickly assess but a human would miss. He said the researches are not 100% certain what the AI is looking at, but the self-taught algorithm did come up with accurate findings when it was applied to new patients.

"That is actually getting right to the core of this study. Essentially [the AI] is a 'black box,' because we just feed the X-ray into the network and it spits out a risk prediction of how likely a patient is to experience some cardiovascular event in the future. But we can't tell what features are responsible for the prediction," Weiss explained. "We can't pinpoint which anatomical alterations or other changes in the image actually contribute to the final prediction."

This type of "black box" AI technology where at AI finds complex patterns based on thousands or hundreds of thousands of tiny image data points may be the way of the future for all sorts of AI-based risk prediction algorithms. If accepted by the FDA, these types of AI will assess all X-rays taken at a hospital or clinic to automatically screen patients in the background and add additional value to every exam. This may aid earlier disease detection and prevention efforts before patients develop an acute condition.

Weiss said the American College of Cardiology (ACC) and the American Heart Association (AHA) both recommend risk assessments for patients to aid primary cardiovascular prevention efforts. He said these assessments are based on nine datapoints to make a prediction. 

"We compared our chest X-ray based cardiovascular risk prediction to this guideline recommended risk prediction, and what we found in this testing cohort of 11,000 patients that we are not significantly worse that the guideline-recommended risk predictors," he said. "Overall, we are as accurate as the current risk predictors." 

Read more details from this study