Digital Twins in an Expansive Future

By Allison Proffitt 

May 2, 2024 | The practice of medicine has struggled to adopt data-driven tools, Dr. Caroline Chung told the audience at last month’s Bio-IT World Conference & Expo. Even when tools are developed, getting them into clinical practice has been slow.  

Chung has unique insight into this problem. She is the Vice President and Chief Data Office and Director of Data Science Development and Implementation of the Institute for Data Science in Oncology at MD Anderson Cancer Center. She is a clinician-scientist and associate professor in Radiation Oncology and Diagnostic Imaging with a clinical practice focused on CNS malignancies and a computational imaging lab focused on quantitative imaging and modeling to detect and characterize tumors and toxicities of treatment to enable personalized cancer treatment. 

She knows that healthcare and life sciences are generating a lot of data, but those data are not necessarily useful and are not optimized for clinical use.  

“When you talk about the explosion of data, there’s certainly a lot of data that we’re generating. How much of that is truly useful to making a clinical decision and improving the care of our patients? How do we enable useful data, both in the development as well as implementation of many of the tools that we’re all working on today?”  

Chung explored the parameters of “useful” data. “Here is the reality check: the availability of useful data in healthcare is somewhat sparse, and so data generation and flow is messy and labor intensive in healthcare today.” And yet the data must be gathered, even if doctors begin to feel like assembly workers as they plug content into EHRs, expanding the volume of data we have. But for models to work, those data must be accurate. Chung warned against “Control V” issues, where errors are copied and pasted throughout the patient record, a problem that is very hard to remedy. On a larger scale, our data are biased toward positive results, as those are the ones that get published.  

All of these are tricky problems, but Chung encouraged data generation with an eventual digital twin in mind. The goal is not simply to add “more widgets into what we do, while doing things the same,” she challenged. Thus far, we’ve been very good at collectively gathering published data and coming up with general trends and recommendations at the population level, she said—not unlike a Farmer’s Almanac, which makes annual predictions based on historic data. These trend data are important and helpful but are not personalized predictions.  

Twins, Defined 

A personal data model might look like a digital twin, though Chung takes issue with the overuse of the term. She participated in creating last year’s National Academies report on Foundational Research Gaps and Future Directions for Digital Twins and she prefers the definition on which the committee settled. “A digital twin is a set of virtual information constructs that mimics the structure, context, and behavior of a natural, engineered, or social system (or system-of-systems), is dynamically updated with data from its physical twin, has a predictive capability, and informs decision that realize value. The bidirectional interaction between the virtual and the physical is central to the digital twin.”  

Digital twins are more than just simulation and modeling, more than an avatar, Chung emphasized. The bidirectional interaction between the digital twin and the physical reality means data moves back and forth. “Unless there’s that bidirectional update, that is not truly a digital twin.” But those data don’t need to be in real time. Digital twins should be fit for purpose, Chung said, and in some cases that may mean updating them on a timescale of days, weeks, or months depending on their use case.  

Considering an appropriate timescale applies to not only how often the digital twin is updated with new physical data, but the frequency with which the model itself is validated. “At what frequency and how are we going to manage a constant updated model and verifying and validating and generating uncertainty quantification that’s reliable?” Chung asked. “This is a major gap that we have.”  

As we improve models, though, there is a risk that we could take the digital twin almost too literally. “There’s a lot of really amazing visuals and imagery that can help convey information,” Chung said, “But how do we start to also convey some of the uncertainty in the images that are being presented so that we can make sound decisions knowing that there is some uncertainty in some of this information coming at us?”  

Those pretty pictures have another downside as well. “If we get bedazzled by the technology that is showing up amazing visualizations and all very exciting,” she warned, “we may end up shortchanging ourselves out of a full potential of a digital twin.”  

Beyond Medicine 

Digital twins hold so much more promise for biomedicine and healthcare than just virtual humans, Chung said. She listed virtual twins of hospital organizations, manufacturing processes, phases of therapeutic discovery and development, and clinical trials. For instance, Chung challenged the audience to consider whether animal models are necessary if we can create digital twins for target discovery. For clinical trials, for instance, digital twins could help us more accurately look at drug delivery cadence than many iterative trials.  

The National Academies’ report also looks beyond virtual humans, covering climate science, engineering, and biomedicine. The sponsoring groups involved in putting the report together include the National Science Foundation, the Department of Defense, the Department of Energy and NIH. Within the report, the authors call for further input from NASA, NOAA, and NIST as well. That breadth is intentional and valuable, Chung argued. Different languages from various disciplines must be harmonized so that we make progress.  

“Imagine a world, let’s say decades ahead, that we have digital twins of the planet; we have digital twins for climate; we have digital twins for health—and they don’t actually interact. We haven’t made them interoperable. We’ve lost a major opportunity.” 

Chung isn’t arguing for completely open sharing, but instead for transparency and enough metadata so models can be cross-calibrated.  

“We want to make sure that we’re building to know that we can actually integrate these aspects so we can actually leverage the full power of the assembly of digital twins moving forward.”