A $4 million grant from the National Institutes of Health is helping faculty researchers in the University of Virginia School of Engineering and Applied Science develop an artificial intelligence tool to personalize treatments for people with heart failure.
The award’s recipient, Miaomiao Zhang , an assistant professor of electrical and computer engineering and computer science, is collaborating with colleagues in biomedical engineering and clinicians at UVA Health to develop GenStrain, an AI model that could one day help doctors determine which type of pacing therapy is most likely to benefit an individual patient.
If successful, GenStrain could turn a routine MRI scan available at most hospitals into a detailed picture of how a patient’s heart muscle moves — helping doctors choose the pacing therapy most likely to work and leading to better cardiac care and patient outcomes.
More than 6 million Americans are living with heart failure, according to the American Heart Association. In many people with heart failure, the heart’s lower chambers no longer contract in sync, weakening each heartbeat.
Implanted pacemaker devices that restore that coordination have improved and saved countless lives, but 30% to 40% of patients receiving cardiac resynchronization therapy, or CRT — the most common of these pacing treatments — still don’t respond as hoped. Zhang and her team hope their approach will help solve that.
Clinicians, including Zhang’s co-investigator, UVA Health cardiologist Kenneth Bilchick, currently rely on two main pacing approaches for mitigating heart failure. Traditional CRT uses biventricular pacing, which stimulates both lower chambers of the heart. A newer option, left bundle branch area pacing, or LBBAP, works through part of the heart’s natural electrical system.
Both approaches can be effective, but doctors don’t yet have a reliable way to decide which is best for a given patient — and that starts with data.
One way to improve those decisions is to measure regional myocardial strain, or how well the heart muscle stretches and contracts with each beat. While there is no clinically guided rule for choosing one approach over the other, strain data can offer clues to which treatment patients are likely to respond to by showing where and how the heart is out of sync.
“Currently, the gold standard for measuring this is the DENSE MRI technique, but it requires extra scan time and isn’t available on most hospital scanners,” said Rishov Paul, a Ph.D. student who has worked on Zhang’s research team for three years.
DENSE, short for displacement encoding with stimulated echoes, precisely tracks how heart tissue moves, but this type of MRI is usually only available in research hospitals and specialized cardiac facilities.
Standard cardiac cine (moving) MRI scans are widely available but can’t capture the same level of detail. That leaves doctors with a dilemma: The MRI scan that best measures strain, and can help them decide which course of treatment is best for heart failure patients, is one most of those patients can’t easily get.
Zhang’s team is developing a generative AI model that learns from DENSE MRI scans what accurate heart motion looks like, then applies that knowledge to standard cine MRI scans, said Paul, the lead developer behind GenStrain.
“The goal is to produce DENSE-quality motion and strain estimates from a standard, routinely acquired cine MRI alone, with no extra scan,” he said.
Zhang began researching how to get more from routine cardiac MRI scans in 2023, when she received an NIH R21 Trailblazer Award . That work laid the groundwork for the new award, which tackles the same problem: helping clinicians match heart failure patients with the pacing therapy most likely to benefit them.
What has changed since 2023 is the project’s scale. The new grant expands Zhang’s access to data from several institutions beyond UVA. Her team is also developing more advanced AI methods designed to work reliably across varied clinical settings.
“We’re no longer focused only on whether the model works at UVA,” Zhang said. “We want to know whether it remains reliable across multiple clinical centers and challenging patient cases.”
For several years, Zhang has worked with Frederick Epstein , UVA’s Mac Wade Professor of Biomedical Engineering and professor of radiology and medical imaging, and Bilchick.
Epstein and his research group have been at the forefront of developing cine DENSE MRI over the past two decades. His team provides the high-quality imaging data used by Zhang’s group to build and evaluate the AI models, while Bilchick helps identify the clinical questions their technology will ultimately answer.
“Our team brings together complementary expertise,” Zhang said. “Imaging scientists provide the high-quality data, engineers develop the AI tools and cardiologists help ensure the work addresses real clinical needs.”
Researchers are collaborating with teams at Virginia Commonwealth University, Baylor College of Medicine, the University of Utah and other partner institutions to test GenStrain beyond UVA. The teams will use anonymized imaging and clinical data to evaluate how accurately the model performs across different patient populations and clinical environments.
First-year Ph.D. student Zafer Yildiz is excited about the model’s potential to help patients.
“I’m interested in building AI systems that are not only accurate, but also reliable and practical in clinical settings,” Yildiz said, but noted that working with medical imaging data is inherently challenging.
Sometimes the research team, which includes undergraduate students, finds inconsistencies between GenStrain’s data and actual anonymized patient records that must be resolved and accounted for in the model.
“Medical scans are not always perfect, and sometimes the details are hard to see even for an expert radiologist,” Yildiz said. “Teaching a model to find them is even harder.”
While AI’s rapid advancement has expanded what Zhang and her team can do, accuracy alone is not enough. The researchers are also developing methods to train their models to explain the underlying “thinking” so that the recommendations are more reliable and interpretable for clinicians.
“For healthcare, AI cannot simply be a black box,” Zhang said. “Clinicians need to understand why the tool is making a particular recommendation.”
Over the next five years, the team will develop and validate GenStrain using data from UVA and its partner institutions. A clinical trial would be a future step, after the researchers establish that the technology performs accurately and reliably in varied settings.
Scott Acton , chair of the Charles L. Brown Department of Electrical and Computer Engineering and American Telephone and Telegraph Company Professor of Engineering, said UVA’s support of interdisciplinary research provides fertile ground for innovative work like Zhang’s.
“This NIH-funded project positions Miaomiao Zhang and her collaborators at the forefront of the transformation of cardiovascular medicine through artificial intelligence and computer vision,” he said. “Their work is helping shape the future of how these technologies are integrated into cardiovascular care and shows the creativity researchers bring to solving these challenges.”
It’s a prime example of the good AI can do if properly applied, Zhang said.
“AI is great if you utilize it in the correct way. It can definitely change people’s lives, especially in healthcare,” she said. “On the other hand, we also need to be carefully developing those techniques so that it benefits society without unintended harms.”
That makes the loop from engineering to clinical application back to engineering all the more critical, because developing AI tools for healthcare the right way is challenging, Zhang said.
UVA actively encourages that kind of interdisciplinary research, and that makes a difference, she added. “Bringing engineers, physiologists and cardiologists together has been essential to moving this work forward.”