Nearly 20 years ago, Amanda Randles, now a biomedical engineer at Duke University, worked at IBM building a supercomputer. About three years in, she had a realization.
“I didn't want to just build the supercomputers; I wanted to go use them to do something interesting,” said Randles. This motivated her to pursue graduate studies and postdoctoral research, taking her career trajectory toward using supercomputing for medical applications.
Patient-specific aortic flow streamlines simulated with HARVEY illustrate three-dimensional blood flow patterns throughout the aorta and its major branches. Vascular geometry courtesy of the Vascular Model Repository at https://www.vascularmodel.com/.
Ayman Yousef, Aristotle Martin, Randles Lab, Duke University
Over the past few years, Randles and her team have used high-performance computing to develop a tool that simulates patient-specific blood flow, or hemodynamics, based on individual medical scans. The tool—called HARVEY after seventeenth-century surgeon William Harvey who first described the circulatory system—allows researchers to study disease mechanisms.1 Building on this, Randles and her team are exploring incorporating real-time data from wearable devices to create personalized vascular digital twins that could guide clinical decisions.
HARVEY Simulates Circulation in the Body
When Randles started working on the HARVEY project in 2010, she had one goal. “The initial question was just: Can you model patient-specific blood flow accurately, and how much of the body do you need to model?” said Randles.
She and her team wrote the code for a fluid dynamics simulation software that could model blood flow throughout the body based on full-body CT and MRI scans.1 At the time, supercomputers took nearly six hours to simulate the flow of one full heartbeat in coronary arteries, recalled Randles. Today, it takes between 10 to 30 minutes, she said, adding that the journey from six hours to half an hour has been both exciting and challenging.
After building and training HARVEY, Randles and her colleagues tested its utility. They used the model to measure the fractional flow reserve—an index that measures blood pressure differences across a narrowed and normal artery—in 160 patients.2 Comparing these values with those obtained via conventionally used invasive methods revealed that HARVEY could accurately measure the fractional flow reserve.
Encouraged by this, Randles and her team sought to expand the scope of their tool. Given the advances in wearable devices, AI, and high-performance computing, the researchers have been exploring using HARVEY to create digital twins that can dynamically represent blood flow by continuously collecting data.
Visualizing blood flow through the pulmonary arteries with HARVEY enables researchers to observe patient-specific simulation that reveals complex three-dimensional flow patterns across the pulmonary vasculature.
Ayman Yousef, Justen Geddes, Randles Lab, Duke University
To this end, Randles and her team iterated HARVEY to leverage machine learning for real-time hemodynamics prediction.3 “We are using AI, but every piece is still running a physics-based flow simulation,” said Randles. “What we're getting from the wearable is just the inlet condition to our physics simulation.” For instance, the model uses heart rate data continuously obtained through a wearable.
Digital Twins Could Offer Preventative Predictions
Randles noted that doctors currently implant devices in the pulmonary arteries of at-risk patients that identify minor alterations in blood flow. “By checking these small changes, we can identify heart failure before you start having chest pain,” she explained.
However, the implantable devices only capture snapshots at certain times during the day. “The whole idea here is if we actually captured how you're responding to exercise and what you're doing throughout the rest of the day, we could then identify that heart failure even earlier,” she said.
According to Randles, preliminary results have revealed that their vascular digital twin approach matches the results obtained via invasive, implanted sensors. Eventually, she envisions that this framework would not only give real-time data but also predict blood flow changes in response to medical or lifestyle interventions.
During their work, in addition to the computational and data-storage challenge, the researchers have had to be cognizant that one day millions of people who use wearable devices might use their pipeline. “The number of people wearing wearables changes what's an acceptable false positive or false negative,” said Randles.
However, despite the challenges, Randles is excited about the prospects of this project. “There's so much potential [in being] able to identify so many different diseases before you have to go to the doctor, before you're in the hospital,” she said.
- Randles A, et al. Massively parallel models of the human circulatory system. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. 2015;1-11
- Vardhan M, et al. Diagnostic performance of coronary angiography derived computational fractional flow reserve. J Am Heart Assoc. 2024;13(13):e029941.
- Tanade C, Randles A. HarVI: Real-time intervention planning for coronary artery disease using machine learning. Lecture notes in computer science. 2024:48-62
