AI and Cardiology

Hypertrophic cardiomyopathy, or HCM, is a genetic condition in which the heart muscle gradually thickens. Over time, it can lead to cardiac arrhythmias or even sudden death. It affects about 1 in 500 people in the United States; among visitors to Duke’s heart center, the proportion is likely higher.

But, as Duke cardiologist Sreekanth Vemulapilli, MD, points out, HCM can be an elusive problem to detect.

“Historically, HCM is underdiagnosed,” he said. “Some patients don’t feel anything. Some have some shortness of breath, or chest discomfort,” he continued, noting that such symptoms could point to many different possible causes.

In practice, what this means is that if you want to diagnose HCM, it helps to be actively looking for it, using an echocardiogram, which uses ultrasound imaging.

Even so, HCM remains tough to spot. The condition manifests differently in different parts of the heart and varies according to disease progression and severity. A physician might easily miss it, even with an imaging exam, or conclude that the thickened heart muscle is due to another cause, such as hypertension. And taking, reading, and interpreting an echo takes time – sometimes a lot of it.

An echo isn’t a static image of the heart, but a moving picture comprising hundreds or thousands of video frames that capture the heart in motion from multiple perspectives. Further complicating matters, some parts of the heart are harder to image than others, and most doctors are trained on echos taken from regions that are easier to image. What’s more, all these images must be viewed and interpreted by a physician in the echo lab.

These challenges help explain why patients aren’t routinely screened for HCM. Doing so would require a physical exam, followed by an electrocardiogram, followed by an echo - diagnostics that require substantial time, effort, and expense. And that’s not all.

“Really importantly, the other cost is false positives,” said Vemulapalli, who explained that if a physician sees something on imaging that suggests HCM, the next step might be a stress test, a cardiac MRI, or genetic testing for the patient and their family – all of which can add to the patient’s anxiety.

But thanks to new capabilities offered by artificial intelligence, Vemulapalli and his colleague, Duke biostatistics professor and AI expert Ricardo Henao, PhD, are crafting algorithmic tools to overcome the challenge of quickly and accurately diagnosing HCM.

“Our algorithm basically evaluates the images within an echocardiogram, and asks, what is the likelihood that this collection of images is consistent with a patient who has hypertrophic cardiomyopathy?” said Vemulapalli. He described the algorithm as a form of clinical decision support, one that doesn’t replace the physician’s judgement, but rather raises a flag to say, “you might want to take a closer look at this.”

But if it’s so hard for human doctors to identify HCM, how can a computer algorithm be trained to spot it?

“This is the place where I think Duke is uniquely positioned to do this type of work,” explained Vemulpalli. “We have structured reports and associated images for essentially all the patients who’ve had an echocardiogram at Duke since 1995 … almost a million echocardiograms and reports.”

When these resources were linked with data from electronic health records, Vemulapalli, Henao, and their team were able to identify patients with confirmed HCM. At the same time, they created a parallel population of patients, matched on characteristics such as sex, age, or other medical conditions, who did not have HCM — a control group. The next step was to apply machine learning techniques to train an algorithm to recognize patients with HCM, using an AI model called a convolutional neural network that’s particularly well-suited to parsing video data.

“The idea is to build a model that takes all the video information for [an echo] study for a single patient and then tries to predict, probabilistically, what is the likelihood that all the images for a given patient are consistent with hypertrophic cardiomyopathy,” said Henao.

The model was also evaluated against “gold-standard” imaging such as cardiac MRI, which helped them confirm that the model’s robustness, and the results were further compared against genetic information.

So far, so good. But there was another wrinkle: having generated a prediction, how could they relay that information to physicians in a way that was useful and trusted?

For example, a physician might be given a percentage value — a likelihood of 75 or 80 percent — that the patient has HCM. The doctor could then weigh that information in light of their experience, individual judgement, and personal threshold for taking action. But that might not be enough.

“The majority of physicians, even though we’re taught to think that way, don’t think that way,” said Vemulapalli. “We need a ‘yes,’ ‘no,’ or maybe a ‘maybe,’” he continued, noting that the team is developing an approach that combines the likelihood estimate with context that might be helpful in guiding the decision.

Part of that important additional context is what Vemulapalli refers to as a “saliency map,” which includes an image with highlighted areas showing which aspects of a given echo were most influential in creating the algorithm’s ultimate score, thus giving physicians a window into its decision-making process.

Henao pointed out this also streamlines the evaluation process for the clinician, who is spared from having to go back to original echo images, with their multiple perspectives and thousands of video frames.

“You are already pointing to where in the image the information that the model used to make the prediction is,” he said.

Although the HCM algorithm is not yet in clinical use at Duke, the team is working to validate the model across multiple research sites. The Duke team is part of an NIH-funded effort, led by cardiologist and principal investigator Svati H. Shah, MD, HS’01, MHS’05, that is evaluating this question at Duke and at three other sites: Mayo Clinic, the Medical University of South Carolina, and Kaiser Permanente of Northern California. This collaboration is looking at whether including other data – such as electrocardiograms or information from the patient’s electronic health record – can further improve the model’s performance.

The team is also collaborating with a vendor to produce an FDA-vetted version of the algorithm that would be commercially available, not just at Duke but at other cardiology centers — a potential gamechanger for what has otherwise been a difficult condition to diagnose and treat. Ultimately, the approach may also be applied to other clinical specialties where imaging plays a significant role in detection and diagnosis of disease.

“Duke has a proud tradition of measuring everything we do,” said Vemulapalli, noting Duke’s role in creating some of the world’s first cardiovascular databases as early as the 1960s. “It’s part of the culture of cardiology here now — everybody tries to measure what they do and improve on it.”