For people who have a serious structural problem in their heart but no symptoms, there has not been a practical way to detect the disorder.
A researcher at the Yale School of Medicine, however, has developed a way to find the seemingly invisible problem using artificial intelligence.
Dr. Rohan Khera, clinical director of the medical school’s Center for Health Informatics and Analytics, said a simple electrocardiogram can be analyzed using AI, which can detect what a basic reading of an ECG cannot.
Khera’s paper about his research appeared in the July 25 issue of the journal Circulation. The work is centered in the Cardiovascular Data Science Lab, where Khera is the principal investigator.
“Many people, up to one in 20, have structural heart disorder, so their heart function is down, but they don’t know that until they develop symptoms and go to the hospital … or (suffer) adverse effects of a certain sort that trigger health care needs,” Khera said.
The disorder is known as left ventricular systolic dysfunction, which reduces the ability of the heart to pump blood. It develops before symptoms appear and medical attention is required, Khera said. The problem can lead to more than an eightfold increase in heart failure and double the risk of premature death, according to Khera’s study.
There are “very inexpensive treatments” for the dysfunction, he said.
“However, we just don’t have a way to know who had this. The way to diagnose it is actually getting cardiac imaging, which is an ultrasound of the heart, or an MRI,” he said. “But those tests are not something you can do with everyone in the community, so this is not feasible.”
Using AI and deep learning, “which are very exciting domains in medicine, we developed a technology that actually uses ECG data,” Khera said.
ECGs, which register electrical activity in the heart, are routinely done during physical exams. They can now be taken on a wearable device such as an Apple Watch, Khera said. About 100 million ECGs are done in the United States each year, he said.
“We have now identified ways to find signatures of structural heart disorders from these very inexpensive and widely available tests,” Khera said. “And that’s the key observation and we are able to do that with accuracy in the 90 to 95% range.”
The Yale study has been repeated in California, Missouri and Texas, and validated in a Brazilian longitudinal study. “So (we’ve) kind of shown that our approach to identifying those with disease really works across the board,” he said.
“Before, when we were doing echocardiograms, when we were only relying on cardiac imaging, there was not infrastructure in place to do that at scale for the population,” Khera said. “But now with these technologies, you can actually get people without any expense or challenge … a diagnosis that is very treatable.”
ECGs, which measure a number of different electrical signals in the heart, are printed out on paper. “We can use the photos of those ECGs and we use computer vision algorithms,” Khera said.
“These are tools built to infer deep signatures of disease from photos. So we take photos of ECGs and identify areas within an ECG that may actually be suggestive of a diagnosis of low heart function. … It’s a hidden feature inside that photo of an ECG that none of us as clinicians can pick up on.”
While the ECG is used to measure the electrical signals, AI can pick up structural abnormalities in the heart by examining the photo.
“Screening for heart disorders with ECGs is likely going to become a more and more common thing as time goes on,” Khera said. “And it’s nice to see AI actually helping pick up things that humans cannot.”
Ed Stannard can be reached at estannard@courant.com.
