In healthcare, we hear the term “digital twin” used more frequently these days. In a recent conversation with Amanda Randles, Ph.D., director of the Duke Center for Computational and Digital Health Innovation, she explained the broader concept as well as the work her lab is doing.

Randles’ lab at Duke University has developed HARVEY (named after William Harvey, a 17th-century surgeon who is credited with first describing the circulatory system). Her lab describes it as “a cardiovascular digital twin engine designed to simulate patient-specific blood flow and vascular dynamics across the full human vasculature. HARVEY enables image-based, physics-driven modeling of blood flow from large arteries down to microcirculation, at computational scales previously unattainable for biomedical simulation.”

Healthcare Innovation: Could you start by describing the work your lab is doing? 

Randles: Our specific lab is focused on creating large-scale digital twins, where we're integrating the use of high-performance computing with physics-based modeling, AI and a lot of computational fluid dynamics to aid in early diagnostics of disease.

HCI: You’re also the director of the Duke Center for Computational and Digital Health Innovation. Are there other types of digital health innovation projects under way? 

Randles: Yes. We have experts in wearables. We have experts in augmented reality and extended reality. It’s blending different directions in the computational digital health space. 

HCI: Could you talk about the concept of digital twins in healthcare more broadly? Is there a lot of exciting work going on in this space?

Randles: There are a lot of examples. It's definitely early days, and we're seeing a lot of adoption, a lot of excitement around it. You have companies like HeartFlow and CathWorks. There are a lot of companies in this space that are using non-invasive methods to capture what they're focused on, which is fractional flow reserve. That's the metric that doctors use to determine if you need a stent or not. If you have a lesion in the coronary artery, and they're trying to figure out if they should stent it or not — how severe the ischemia is — it’s really based on the pressure gradient across that narrowing. Conventionally, you put a guide wire into the artery and measure the pressure before the lesion and after the lesion, and it's really just the ratio of those two pressures. Now they're using these FDA-approved tools to actually do this non-invasively, using physics-based computational models. They’re making a digital twin of the patient, running a blood flow simulation in that digital twin, and then measuring that fractional flow reserve in the digital twin instead of in the patient. 

HCI: What does it take to create the digital twin of the patient? Imaging?

Randles: The imaging is important. Everybody's anatomy is so different that you really need tailored anatomy. Every tool has a different way of doing it. There are some that go from MRI, some that go from CT, and some that are going from conventional coronary angiograms. But you need some way of getting that 3D anatomy simulation. From there, every tool is a slightly different version of setting the boundary conditions for your physics model. The tools are running physics-based flow simulations.

HCI: Could you talk about the development of HARVEY?

Randles: We have been working on HARVEY since 2009 or 2010. It has evolved over time. Initially, it was very much in line with this kind of fractional flow reserve idea. Back in 2009, running those flow simulations would take the world's biggest supercomputers. Our 2010 simulation took the entirety of the world's biggest supercomputer, and then it took six hours to run one heartbeat.

The goal has been to run high-resolution simulations that are much longer. We’re running three-dimensional fluid dynamic simulations. Initially we wanted to just get a heartbeat at a high enough resolution that you could do something useful. We've spent the last 15 to 20 years trying to make it faster and not require the whole supercomputer and to run it in the cloud. We're also using it now to connect to wearable devices, so we can get not just one heartbeat, but drive the flow simulations and capture 3D flow models over longer periods of time. HARVEY is really the engine for the physics simulation of how you do the computational fluid dynamics. 

HCI: From a clinician's viewpoint of the value of this, is it the same use case you were describing — trying to decide whether someone might need a stent or not? Or are there other use cases for cardiologists? 

Randles: Initially we focused a lot on the diagnostic question of do you need a stent or not. But in connecting it to the wearables, we’re trying to identify if we can determine if something's going wrong earlier and do that non-invasively. We’ve done a lot of work lately with heart failure. For heart failure, right now, you have an implantable sensor that is measuring your pulmonary artery pressure. We've been comparing HARVEY with those results to see if we can get that pulmonary artery pressure non-invasively. Those sensors can only measure it once a day while you're lying down, so you're missing things like how are you responding to exercise? What's your heart recovery? You're missing a lot of that dynamic data. So we're really pushing to try to get a more complete picture of the patient. 
We've also done a lot of studies to go beyond the heart. We've looked at cerebral vasculature and aneurysm risk. Anywhere you have large vessels where you may have a narrowing, we're broadening to other areas of the body as well.

HCI: Are the cardiologists and other clinicians receptive to this? Does it take a lot of convincing or explaining that this is could be better in some cases than what they're used to doing as a gold standard of care?

Randles: They’re super supportive. The cardiology field has been one of the more forward-looking and open to this kind of research. HeartFlow really set the stage that this can be useful. 

We've been doing a lot of studies to look at how we can get that data back to the cardiologists in a way that's useful. We've done a lot of work combining HARVEY with extended reality and augmented reality interactions. A lot of those studies have been done with the cardiology department here at Duke. When we run these user studies, it's very hard to get time with the doctors because they're busy, but they are so excited by this that they will spend hours with us, playing with the virtual reality and what they can do with it.

HCI: I read that HARVEY could also be extended to cancer cells and what drives disease development there…

Randles: One part of our lab is looking at cell-based mechanics. We can model deformable red blood cells. We have cancer cells, red blood cells, and then we can also handle adhesion. We go down to the fine scale of individual ligand receptor pairings. We can model the cancer cell moving through the body, and then actually capture individual ligand receptor bonds as they're forming and see how those interactions are affecting the cancer cell, how long it’s spending at different locations in the body, and how the forces are interacting with it. Because we've been focused on large-scale computing, we can model hundreds of millions of red blood cells around that cancer cell and really see how it's interacting in the body, with realistic geometries. One question is: Can we understand what it is about the cancer cell that's causing it to spend more time at different places in the wall? The goal is to try to find new therapeutic targets. 

HCI: So does that involve partnerships with oncology researchers, too?

Randles: Yes. And with bioengineering and mechanical engineers. We're collaborating with labs that are bio-printing different microchips that we can then run the cancer cell experiments through, and make sure we're really capturing the right properties about that cancer cell. 

HCI: We are writing about this huge proliferation of AI-related innovations in the clinical space involving large language models. Is AI also impacting this kind of research? 

Randles: We're using AI a lot, but it is slightly different. We're informing AI models, and we are using AI to analyze the results of the big simulations in trying to understand: What are these biomarkers? For instance, we know that pulmonary artery pressure changes a few weeks before you go into heart failure; it’s a predictive, it's diagnostic. It can help us identify it. But are there biomarkers that change six weeks beforehand? That involves combing through petabytes of data about every individual person trying to find that biomarker. An AI surrogate that can be deployed at the edge is much more computationally efficient.

HCI: Do you think that the concept of digital twins will become much more prevalent, and that our readers who work in healthcare will become more familiar with it soon? 

Randles: I think that's 100% where we're going, and it's not 20 years away, right? I think that in the next few years we're going to see these be much more prevalent. One of the big innovations we've had lately is we have a new algorithm that lets us not just model a heartbeat, but we've worked on six weeks of time. This week, we're going to try to run our first simulation to run a whole year of someone's 3D blood flow. 

We're shifting and using these new algorithms to get to much longer time periods. The reason this is important is because we now have the wearable devices to get that data. Years ago, when those weren't as ubiquitous, we didn't need to go beyond a few heartbeats, because you never had the input to really try. This opens that up. With so many people using wearable devices, you have access to your continuous data as you're going about your daily life. A lot of these digital twins can now make use of all of this data that we're getting. That’s going to be the big pivot, where we finally have all this data and we have all these advances in AI, so now we can actually integrate all this multimodal data, and we're kind of at that precipice where we can do something with it.