Science Friday - Could a ‘digital twin’ help you get better health care?

Episode Date: March 17, 2026

There’s an idea bubbling up in medicine called the “digital twin.” The concept is to take personal health data like genetics, blood test results, tissue samples, MRI scans, and family history, and create a digital model of a patient that can be used to predict how a treatment might work for them. Think personalized medicine supercharged by AI.  For example, cancer researchers are working on models that would create radiation and chemotherapy treatment plans based on the specifics of a patient’s tumor. But these models aren’t ready for the clinic yet, and with so much patient data involved, privacy concerns abound.  Host Flora Lichtman talks with Caroline Chung, a radiation oncologist at the forefront of digital twin research. Guest: Dr. Caroline Chung is a radiation oncologist and the co-director of the Institute for Data Science Oncology at UT MD Anderson Cancer Center. Transcripts for each episode are available within 1-3 days at sciencefriday.com. Subscribe to this podcast. Plus, to stay updated on all things science, sign up for Science Friday's newsletters.

Hey, I'm Flora Lickman, and you're listening to Science Friday. There's an idea bubbling up in medicine. It's called digital twins. If it's giving you second life vibes, that's not it. The rough idea is to create a digital model of you. So taking your personal health data like genetic info, blood test results, imaging scans, family history, and compiling it into a simulation that can be then used to help predict how a treatment might work for you. think personalized medicine supercharged by AI.

So how would this be useful? How would it work? What are the concerns? Here to tell us more is Dr. Caroline Chung, a radiation oncologist at the forefront of this digital twin research. She's co-director of the Institute for Data Science Oncology at UTMD Anderson Cancer Center. Caroline, thanks for being here. Thanks for having me. Did we get that description of digital twins right?

I think that that's a good highlight. level overview of how digital twins potentially could be used in the future. But. Having said that the definition of digital twins has been blurry at best, and many people are using this term to mean many different things. Some people imagine a digital twin to be their digital avatar or some visual representation of all of the data that we're collecting of you within the medical space. But these are not necessarily all digital twins from a true perspective of what a digital twin was intended to be as it was born out in the aerospace engineering industry.

And so a digital twin is more than just a model that can actually predict what is going to happen to you. It really is an ongoing interaction between what the model will predict, your actions based on what information you receive, and continued data collection to update the information and predictions using that model. So it really is something that journeys with you through time. You mentioned aerospace. Where did this idea for digital twins come from? So digital twins were born out of the fact that you probably don't want a plane losing its wing before you realize you should have repaired something.

So the idea was that if you can model the different, parts of a plane or any sort of spacecraft. It's a very costly endeavor to actually design a bunch of models that will actually simulate all of the different possibilities of what could happen to something that is being put up in flight or put up into space is, A, very costly, but B is very difficult to actually experiment with. And so what happened was they created a digital replica of these devices and these vehicles so that you could actually pressure test in a digital space,

all the possible things that could happen. And then they took the most promising designs and moved them forward into models that they would actually build out and test in the physical world. And that bridging between the physical and the digital world, that continuous bridging is ultimately the design of a digital world. twin. Applying that to biology is a whole other level of complexity because, as you can imagine, mechanical parts that humans are designing, it's somewhat easier to anticipate what will happen with physical laws in place, and a lot of the physical laws are known. Whereas the biology space,

we are still discovering a lot of the molecular mechanisms even today. And so there's a lot of lot of gaps and knowledge that we're working with as we try to build out these digital twins in healthcare. Yeah, that was going to be one of my questions. I mean, do we have the data to actually underpin these models yet? Yeah. And I think that we do have the data to design some of these models with confidence because we understand the mechanics of, for instance, a beating heart and the muscle contractions and being able to measure that muscle contraction, we have capabilities to do that, we can model heart flow. There are general physical laws and equations around fluid dynamics, for instance. And so there are certain areas where this is readily positioned for digital twins and

other areas where there's a lot more gaps and it would take a lot more work for us to fill in those gaps as we to realize the full potential of these digital twins. For your work, are you using an AI engine like clod or something to power this digital twin? So the digital twins are not necessarily all using AI models. Many of them are using physics-informed or what we call mechanistic models. Others are using hybrid models combining AI with the power of the physics-informed models. And the difference is that many of these mechanistic or physics-informed models can put specific constraints so that you're not getting completely wacky results when you actually put the

data in and give you your predictions. You're an oncologist. What problem would a digital twin, you know, potentially solve for you? How would it make it easier to do your job or to do it better? So one example of a digital twin that we're actually working on is as a radiation oncologist, one of the challenges is defining what sub-region of the tumor needs higher doses of radiation, what parts of the normal tissues could we spare and reduce radiation dose while still effectively treating the tumor? And those are daily questions. And so if we've started to design digital twins that can anticipate where the more aggressive tumor cells or resistant tumor cells are so that we can actually adapt and personalize the radiation treatment for each individual tumor

that we're treating and each individual patient that we're treating and the tumor in that patient. And so accounting for the toxicity as well as the effectiveness of the radiation, we can actually start to modify that radiation potentially very early on through the treatment so that we can really make an impact on the outcome. Have you tried this with patients yet? There have been some designs of studies where we've looked at differential dosing of radiation. My colleague, Kiko Enderling, who is a leader within the Institute for Data Science and Oncology, has run clinical trials with his clinical partners.

We are also designing clinical trials proactively at MD Anderson around use of digital twins moving forward. Give me some other examples where you think digital twins might be really helpful in medicine. So in other examples, there have been digital twins built in the cardiology space, and they've mimicked and looked at flow of blood through the cardiac vessels in terms of anticipating whether someone needs a catheterization before a heart attack happens. That would be a very useful application. In cancer, we've actually explored other applications, such as looking at whether we could look at the imaging and all the risk factors that someone has to personalize their follow-up of screening tests. So screening tests may need to be started earlier in some

individuals, may not need to be done annually in everybody. And so we can start to balance out both alleviating anxiety of going through a screening test and really personalizing to that individual saying, you don't need to come back for another three years. And so could we, we actually really bring people back earlier if they're at higher risk and leaving people to their regular day lives and staying away from getting another test if it's not necessary. And so this is both cost effective as well as helping the patient. Another example is we've actually simulated giving the same chemotherapy with different schedules to get better results, which is a really interesting finding that we found through the

simulations is that you can take the same total dose of standard chemotherapy for ladies with breast cancer and get very different outcomes if you were to deliver it with the different schedule than what the current standard of care where everyone gets the same schedule of chemo. And if you actually personalize it to that individual person's physiology and tumor biology, you may get improved outcomes. And it's not necessarily with additional drug costs, which is very interesting. I mean, we're talking about sort of individual systems. Is the goal to create a digital twin that can put it all together,

that can replicate every system in a person's body and then compile it? So there are efforts globally. There's a virtual human global symposium that has been bringing together experts from around the world, working on, as you say, the individual system level digital twins. So there is an aspiration to build up, build out a digital twin of an entire human. Having said that a digital twin really needs to be fit for purpose. Because in order for us to make the investment, if you think about having all of this data flow through all of these models that are giving all of these predictions, what are we actually using it for? And it could be that we're using it for scientific discovery,

understanding the physiology and human body better. And that could be one goal, in which case it may be worth the investment to actually bring all of these pieces together. But it may be that we're trying to figure out a very specific clinical question that we're asking, in which case we probably don't want to be connecting, or we probably don't need to connect all of the different pieces together to answer that very specific question. Right. Also, like more ways for it to go wrong. Right. Exactly. And the final piece around this is that once you start to build together all of the different pieces into one entire human being, the privacy

issues do escalate. And so it becomes very identifiable if you have all of the different pieces coming together because there's probably only one of you that have all of those parameters in a digital twin. Yes, I want to talk about this. Don't go away because there is a lot more to discuss on this, including these exact concerns, concerns we might have about privacy and ownership. Don't go away. Let's talk about privacy. I mean, first of all, if your doctor creates a digital twin of you, who owns it? It's a very good question. And it's also a fact that, that even if we were to say that the person owns it, do they have the capability to access it without the supports of all of the systems? If you think about all the data flows,

the model supports that need to be in place for that digital twin to stay alive, it becomes a very complex scenario. Like even being able to open it and look at it might not be possible unless you have the right software or whatever. Yes, yes. So the sooner we sort these pieces out, the sooner we will be ready to even bring forward digital twins into the healthcare space because it is a very sensitive issue. What happens if a prediction that the digital twin ends up being wrong?

Like the patient has a bad outcome? Who's culpable? Yes. So I think this is true of any model. I think a digital twin is another frame of information, and it's a way to bring the information together. And ultimately, as a practicing clinician, I consider the clinical decisions made collectively with my patient.

And I think that there are many things that get weighed in, not just the pure measurements that are coming off systems that feed into a digital twin, but there are other aspects around personal, social pieces that will contribute to the overall decision. And I think that those things need to be considered as we embrace technology. So we can embrace technology but maintain the humanity of medicine. You know, one thing that we've covered on the show is about how we humans have this bias to think that, computers are right, that algorithms are right, that AI is right, and we see it in other spheres. Do you think about that, like how it changes the practice of medicine? If you have a digital

twin suggesting to do some certain thing, even if it goes against your intuition as a doctor, like how do you weigh that or how do you think about those decisions? Yes. And in fact, this is something that's very interesting. at this point in time, in this moment in time, as we see AI coming in to play, not even digital twins, but just this kind of technology coming at us from all directions.

There is an emerging science where we need to learn more about this human machine interface and how it's actually affecting our cognitive processing and our decision-making. I would argue that there's probably two, camps, those who are very enthusiastic and have this bias towards thinking the machine is always right, as well as those who are very skeptical about these models and they think there's the human cognitive bias thinking that the humans still know better. And I would hope that we get to

a place where we figure out how to balance those pieces where we're critical of the outputs at the same time open to suggestions that may not make sense within our own knowledge framework. And so understanding how to have that information presented and even that user interface and what information in what pace and what way will actually facilitate that critical thinking is a really important science that we need to figure out. Do you fall into one camp or the other? I'm in the camp that we need to figure out the better way to present. the information to create pause.

I think that one of the pieces that has come amidst the whole AI era is everyone wants things faster. There's a consumer mentality and wanting to get to an answer faster. And then because speed and acceleration is perhaps an innate goal with this technology or what's been marketed to us as the goal, there is less pause that happens. And the pause is probably the critical piece that will allow us to be discerning being able to think through one of the pros and cons of every decision that we make. Caroline, do you think that these are inevitable?

Like, how far off are these? Or is it still unclear whether they will become a part of mainstream medicine? I would say that there are emerging versions of digital twins that are coming. So I think that if you ask if some forms of digital twins will come to fruition at some point in time, it's likely that it will. There are a number of pieces that we need to address for this to systematically and systematically be adopted across all practices and making sure that there is accessibility for all.

And that comes down to how are we actually generating data in medicine? How is it actually flowing? And so addressing some of these operational, legal, regulatory practices are necessary, hoops that we'll have to jump through to make these readily available to everyone across medicine. Dr. Caroline Chung, Radiation Oncologist and co-director of the Institute for Data Science Oncology at UTMD Anderson Cancer Center. This episode was produced by Shoshana Bucksbaum. If you want to help us tweak our model to better serve you, please input some data. 8774-4-Syphry is our number. We are always looking for your story ideas, your questions, your comments,

and concerns. Give us a ring, 8774 Sigh Fry. I'm Flora Lichten. Catch you next time.