5
Digital Twins

Caroline Chung, vice president and chief data and analytics officer, co-director of the Institute for Data Science and Oncology, and professor of radiation oncology and diagnostic imaging at the University of Texas MD Anderson Cancer Center; and Sylvain Costes, former National Aeronautics and Space Administration data officer for space, biological, and physical sciences, moderated a session devoted to understanding digital twins, their potential, and the challenges to reaching that potential. Four speakers offered presentations, followed by a discussion period.

DIGITAL TWINS AND ARTIFICIAL INTELLIGENCE FOR PRECISION MEDICINE

Jun Deng, professor of therapeutic radiology at the Yale School of Medicine, spoke about the use of digital twins and artificial intelligence (AI) in precision medicine.

Deng began by defining precision medicine as using information about a person’s genes, proteins, environment, and lifestyle to prevent, diagnose, or treat disease—often described as targeting “the right patients with the right treatments at the right time.” AI may be able to revolutionize precision medicine through early disease detection, real-time patient monitoring, drug target identification, causal gene identification, and management of phenotypic and genetic heterogeneity. Modern medicine, Deng emphasized, is increasingly becoming a science of information.

He continued that digital twins represent a paradigm shift in healthcare technology. Deng defined a digital twin as “a digital representation of a real-world physical system or process that serves as the effectively indistinguishable digital counterpart of the original for practical purposes such as simulation, integration, testing, monitoring, and maintenance.” The critical aspect is real-time synchronization—that is, the physical and virtual systems are linked to exchange information continuously.

A 2024 National Academies report, Foundational Research Gaps and Future Directions for Digital Twins, illustrates how digital twins function in medicine (NASEM, 2024). Data collected from patients through imaging, laboratory tests, clinical assessments, tissue samples, and biosensors inform models that simulate what happens in the individual’s body. The digital twin relies on both individual patient data and general models of human body operations, creating an ongoing cycle where predictions inform treatments and treatment results feed back into the digital twin. Deng argued that verification, validation, and uncertainty quantification (VVUQ), which is the systematic approach to ensure that computational models are mathematically correct,

accurately represent real-world phenomena, and properly characterize confidence levels and potential errors in predictions, are essential components.

A digital twin transcends simple computational models or AI algorithms. While these technologies contribute to achieving digital twins, they do not constitute the totality. He noted that digital twins involve three components: a physical entity, a virtual replica, and a connection between the two. Deng stated that they should embody the “five I’s”: individualized, interconnected, interactive, informative, and impactful. He stated that, currently, the examples of healthcare applications include personalized medicine, clinical trials, biomarker and drug discovery, biomanufacturing, device design, surgical planning, hospital management, and personal wellness.

Digital twins in radiation oncology involve multimodal patient data, multiscale modeling, and high-performance computing. Benefits include enhanced ability to predict patient outcomes, optimize treatment plans, and develop innovative research tools. However, barriers include the need for centralized data commons, patient-specific data assembly, and multiscale modeling (Jensen and Deng, 2023).

Deng outlined five potential applications in radiation therapy: real-time monitoring and comprehensive analysis of cancer patients’ health status; treatment plan optimization based on individual patient simulations considering anatomy, treatment history, toxicity, goals, and preferences; virtual clinical trials; predictive toxicity modeling; and proactive maintenance with automated quality assurance.

For building digital twins, Deng proposed a blueprint based on five questions: What are the goals? What level of complexity is needed? Are required data available or accessible? Can uncertainty be characterized? Does a proper mechanistic model exist?

Deng then described three specific applications his team developed. First, the team modeled time-series patient data on metabolic indices collected before, during, and after cancer treatment. Since differences in these time-series data correlate with treatment outcomes and quality of life, understanding temporal correlations enables prediction. This model (Hou et al., 2022) successfully predicted future metabolic indices, although noisy data affected forecasts. To address this, he said, the team worked in latent variable space rather than data space, creating latent ordinary differential equations to solve variable correlations over time before mapping back to data space, to achieve “much better prediction into the future.”

Second, the team created patient-specific digital twins following virtual radiotherapy workflows, enabling tissue characterization, radiation therapy assessment and adaptation, and monitoring. With corresponding health prediction and treatment outcome trajectory models, physicians can present various options to patients for shared decision making. The team also developed whole-body, all-organ toxicity monitoring, with a prototype dashboard that warns of potential toxicity and prevents adverse effects.

Third, the team developed biology-guided radiotherapy using positron emission tomography (PET) monitoring. PET signals reflect tumor metabolic activity, and with multiple scans before and during treatment, they accumulate time-series data. By integrating mechanistic modeling with patient-specific longitudinal PET scans, the team created individualized digital twins to predict non-small-cell lung cancer responses to radiation therapy.

Deng discussed four fundamental challenges in developing individualized digital twins. Data acquisition, integration, standards, and quality pose difficulties for digital twins and AI applications. Creating effective multiscale modeling amid human biology’s complexity proves challenging, he said, particularly at the microscopic level where “there are so many biological phenomena we do not know quite well.” Ensuring responsible AI applications—fair, transparent, accountable, robust, safe, private, and secure—represents another hurdle. Finally, he stated, developing necessary computing infrastructure is important.

The overarching challenge in developing digital twins involves fusing multimodal data from functional and molecular imaging, radiomics, liquid biopsies, whole-slide pathology, genomic profiling, single-cell profiling,

patient-derived cell assays, intravital imaging, fitness trackers, and implantable sensors. As Deng noted, “Each technology gives us new light on patient health state, but it is a challenge to coherently fuse them together.”

Looking ahead, Deng envisions human digital twins becoming fully functional and autonomous, deployed as human digital agents serving as true health companions and safeguards. These agents could interact with humans in the physical world and with other agents virtually. “Imagine that millions of those digital twins interconnect, and they just automatically . . . try to do some new knowledge discovery and correlation,” he said. “That would be very powerful.”

In his final remarks, Deng noted that human digital twins hold tremendous potential in precision medicine. Success likely involves deep understanding of human biology, with multimodal data integration and multiscale modeling as key challenges to overcome. Cross-disciplinary collaborations will likely be important as well.

MULTISCALE DIGITAL TWINS FOR PERSONALIZED RADIOPHARMACEUTICAL THERAPY

Greeshma Agasthya, assistant professor in the Nuclear and Radiological Engineering and Medical Physics Program at the Georgia Institute of Technology, spoke about using multiscale digital twins to personalize radiopharmaceutical therapy (RPT).

Agasthya began by highlighting the transformative potential of RPT. “The big promise of RPT,” she explained, “is that it can selectively target cancer cells while it spares the normal tissue.” This selectivity is achieved by attaching radioisotopes to molecules that target specific proteins or ligands on cancer cell surfaces and then injecting the composite molecule into the bloodstream where radioisotopes concentrate around tumors.

She noted that despite Food and Drug Administration (FDA) approval for late-stage cancer treatment, current dosing remains largely uniform across patients, with only one drug—Xofigo—calibrated according to patient weight. This represents a gap between current practice and RPT’s ultimate goal of personalized dosing to maximize cancer cell death while minimizing unwanted side effects.

To bridge this gap, Agasthya’s group is developing multiscale human digital twins that operate across multiple biological scales simultaneously. The group begins with a whole-body digital twin encompassing all major organs and structures and then introduces digital twins of multicellular tumors to examine radiation effects on subcellular components within those tumors.

Agasthya stated that for RPT applications, digital twins possess four critical capabilities: tracking radioactivity across scales (since some radioactivity reaches off-target organs despite molecular targeting), estimating radiation dose across scales, simulating drug uptake, and predicting patient-specific outcomes.

The development process begins with Duke University’s Center for Virtual Imaging Trials’ virtual human population, created from computed tomography (CT) images. These images are manually segmented to create extended cardiac-torso phantoms—that is, replicas representing both human anatomy and physiology. “We can build these digital twins for any new patient that walks in,” Agasthya explained, and by using CT images from different treatment time points, they create digital twins that evolve over time. The team layers additional complexity by incorporating organ-specific representations, tumor models, and even cellular DNA models to study how three-dimensional DNA conformation affects radiation therapy outcomes.

Agasthya illustrated the team’s approach through a study of lutetium-PSMA (Prostate-Specific Membrane Antigen), an FDA-approved RPT that would be examined at unapproved doses. Starting with a male phantom (body mass index 24) and adding physiologically based pharmacokinetic models, the team determined time-dependent radioactivity concentrations in different organs, particularly peak concentration times. This analysis revealed when the agent would be most effective in target organs and least effective elsewhere, enabling estimation of maximum radiopharmaceutical impact timing.

Agasthya explained that once they determine cumulative organ dose, the team members simulate dose distribution to both targeted and surrounding organs, modeling vascular systems through computational fluid dynamics to examine radiopharmaceutical perfusion and achieve spatiotemporal distribution mapping. At intermediate scales, they use computational models to seed cancer and observe tumor growth or shrinkage responses to therapy, while at the smallest scale they examine DNA damage using Hi-C (genomic analysis technique) data on three-dimensional DNA structure to understand radiation-induced double-strand breaks. By integrating this multiscale approach, they calculate DNA damage, repair mechanics, and predict cell survival fractions for specific radiation doses at particular body positions.

Moving this technology into clinical practice faces multiple hurdles. First, accumulating sufficient information for treatment personalization proves challenging, particularly given RPT’s relative novelty. Additionally, imaging typically occurs only at treatment initiation, not between radiation cycles, and treatment plans change only when clinically observable quantities like creatinine levels shift. While the team ultimately aims to extract more patient information, it currently works with available clinical trial data.

The multiscale nature creates numerous complexities, particularly in multiscale VVUQ. The team works to integrate across temporal and spatial scales while understanding interactions among radiation physics, human biology, and drug uptake biochemistry. Uncertainties at one scale propagate across others, compounding data requirements, computational costs, and patient variability challenges. To address these issues, the team is building cellular digital twins for comparison with live cell experiments to enable VVUQ. The team is also developing validation methods for double-strand breaks.

Future work for Agasthya’s group includes developing patient- and disease-specific multiscale tumor modeling, which is currently lacking. “We are also building AI-driven surrogate models to reduce the computational complexity of these digital twins,” Agasthya noted, while building toward virtual clinical trials of RPT.

Currently, she participates in an Oak Ridge National Laboratory study that created 491 virtual patients with 10,000 simulations of various CT protocols, generating 125 terabytes of data currently under analysis. This massive dataset represents the scale of computational resources for comprehensive multiscale digital twin development.

In her closing remarks, Agasthya stated how her work demonstrates the potential for multiscale digital twins to revolutionize RPT through personalized dosing. By bridging scales from whole-body physiology to DNA structure, these digital twins can optimize treatment efficacy while minimizing side effects. However, significant challenges remain in validation, computational complexity, and clinical integration before this technology can fulfill its transformative potential in cancer treatment.

CARDIAC DIGITAL TWINS: FROM THE ACADEMY TO THE CLINIC

Charles Taylor, director of the Center for Computational Medicine at the University of Texas at Austin, spoke about cardiac digital twins. In 2010 he started a company, HeartFlow, that was focused on the diagnosis of coronary artery disease. Much of his presentation focused on technologies offered by HeartFlow and his story of utilizing CT technology and images to develop this digital twin model.

Coronary artery disease, characterized by atherosclerotic plaque buildup in coronary arteries, represents one of the primary causes of heart disease. While progressive narrowing reduces blood supply to the myocardium and causes angina during exertion, heart attacks typically result from ruptured plaques that form blood clots and cut off myocardial blood supply.

Taylor highlighted the diagnostic difficulties plaguing coronary artery disease assessment. He gave the example that roughly 80 percent of emergency room patients with chest pain do not have coronary artery disease, and among those who do, distinguishing between mild-to-moderate disease (manageable with medical therapy) and severe cases (requiring interventions like stent implantation) remains challenging.

Traditional noninvasive functional tests provide no direct information about coronary artery blood flow or disease. Consequently, many patients with positive functional tests undergo invasive cardiac catheterization only to discover no significant disease, while serious cases are missed due to false negatives. The fundamental problem, Taylor explained, is that these tests do not directly visualize coronary arteries and cannot measure disease at earlier stages. Even with substantial plaque buildup, vessels adapt through compensatory remodeling, maintaining myocardial blood supply until plaque reaches advanced levels.

Current assessment involves cardiac catheterization with pressure wire insertion to measure downstream pressure divided by reference aortic pressure under stress conditions (typically adenosine-induced). This ratio, called fractional flow reserve (FFR), a diagnostic measurement used to assess the severity of narrowing in coronary arteries, guides treatment: values greater than 0.8 indicate medical management sufficiency, while values less than 0.75 require revascularization. However, these invasive tests reveal that 60 percent of patients lack obstructive coronary disease, he said.

Taylor’s solution involved using FFR derived from CTs to build patient-specific coronary artery models to simulate blood flow and predict pressure losses, enabling physicians to “see” arterial conditions without invasive procedures. After years of development, this digital twin technology—FFRCT—now operates in approximately 1,300 U.S. hospitals.

The FFRCT process encompasses three steps: anatomic modeling, physiologic modeling, and functional assessment using computational fluid dynamics. The most challenging aspect involves obtaining precise anatomic representations from imaging data, as coronary artery disease reduces arterial diameter and calcium in plaques creates significant artifacts in CT scans.

To overcome imaging limitations, Taylor’s team collected data from patients who underwent both invasive cardiac catheterization and CT scans, using these datasets to train algorithms that create clearer images from CT scans alone. The team developed a human-in-the-loop AI foundation combining deep learning algorithms with analyst review and correction of centerlines, boundaries, and other features. These corrections feed back into databases training next-generation algorithms, creating continuous improvement cycles that eventually enabled processing of cases human analysts could not interpret.

Feedback mechanisms further enhanced results. Data processing occurred centrally with reports sent back to clinicians, and any discordance between predicted results and measurements triggered investigation and adjustments. As Taylor noted, customer-reported false negative and false positive rates dropped steadily between 2018 and 2023.

The second step utilizes anatomic information plus form-function relationships to infer patient-specific physiology, customizing population-based physiology models for individual patients based on their anatomic data. Finally, computational fluid dynamics quantifies blood flow and pressure in the individual’s digital twin, providing comprehensive flow and pressure data for the entire coronary tree.

FFRCT technology is being adopted clinically for more than 400,000 patients across 1,500 centers, including many top U.S. heart hospitals. The technology receives uniform payment coverage in the U.S. healthcare system, and England’s National Health Service mandates its use.

Multiple studies have validated its diagnostic accuracy, establishing FFRCT as the most accurate noninvasive test for predicting FFR (Driessen et al., 2019; Okonogi et al., 2021). Some studies demonstrate that FFRCT can improve ischemia patient survival rates (Zellans et al., 2021). FFRCT also enables identification of chest pain patients who can defer cardiac catheterization safely and those most likely to benefit from revascularization. The technology serves dual purposes in clinical practice: routine patient assessment and catheterization laboratory prediction and treatment planning.

Taylor emphasized that FFRCT represents the successful transition of digital twins from academic research to clinical reality. This transformation demonstrates how sophisticated computational modeling, enhanced by

AI and human expertise, can address real-world diagnostic challenges while improving patient outcomes and healthcare efficiency.

The FFRCT story is an example of how digital twin technology has the potential to revolutionize medical practice. By creating patient-specific models that simulate physiological processes, clinicians can make more informed decisions without subjecting patients to unnecessary invasive procedures. The technology’s widespread adoption and proven clinical benefits validate the concept that digital twins can bridge the gap between computational innovation and practical healthcare delivery.

Taylor ended by stating that his work demonstrates that successful medical digital twins include not just sophisticated algorithms but also robust validation processes, continuous improvement mechanisms, and seamless integration into existing clinical workflows.

DIGITAL TWINS FOR DISEASE MODELING AND DRUG DEVELOPMENT

Jon Walsh, co-founder and chief scientific officer at Unlearn, spoke about the application of digital twins to clinical trials. He began by noting that in high-energy physics, simulations are a valuable tool for making experiments more sensitive. Physicists simulate everything—the detectors, the background, and the experiment—because the experiments are very expensive and simulations are a way to get the most return from them.

Clinical trials represent an $80 billion-a-year industry with complexity rivaling high-energy physics experiments, making them prime candidates for simulation technology, he said. Walsh explained that most clinical trials for new drugs follow three phases: phase 1 tests safety, phase 2 evaluates efficacy, and phase 3 provides regulatory-grade evidence of safety and effectiveness.

The current standard approach uses randomized controlled trials (RCTs), where participants are randomly assigned to treatment or control groups. While RCTs are considered the gold standard for experimental evidence, they have significant inefficiencies. Patients with diseases enrolling in trials want treatment, not placebos, and the large number of concurrent RCTs with similar control groups creates waste.

Digital twins offer a solution by maximizing utility of RCT data. “We think about using digital twins as adding information about the participants who are already enrolled in a trial,” Walsh explained. This approach helps researchers learn more without adding participants—or achieve the same insights with fewer participants. Digital twins bring lessons from previous clinical trials and observational studies into new trials while maintaining the current RCT’s integrity.

In clinical trials, digital twins function as sophisticated predictive models. Using an example from an amyotrophic lateral sclerosis trial, Walsh described how a digital twin starts with clinical data from patients’ baseline visits before patients receive the trial drug. The goal is to predict what will happen during the study. “This is a basic machine learning problem,” he explained. “You have data about a patient’s health. [They] can be complex. You want to build a probabilistic model of longitudinal health for patients with the disease, and then you want to make comprehensive probabilistic predictions of their future outcomes, and that’s the digital twin.” The twin can be personalized, with future data providing feedback to improve predictions. These predictions inform RCT decision making (Alam et al., 2024).

Building these models involves extensive data collection and curation—a resource- and time-intensive process. “If you do not get it right, then you have a lot of problems building good, robust models [and] you spend a lot of time doing harmonization across different data sources,” Walsh emphasized. Sources include clinical trial data, observational data, and registry data.

After data preparation, models are trained using architectures designed for different disease progression patterns and then validated across multiple performance dimensions. The model’s predictions use Markov processes with Monte Carlo sampling. “An individual patient’s data [are] represented through many different

sample draws of their future outcomes, and you can take the mean, the variance, [and] any other statistics you want,” Walsh explained.

The most regulator-acceptable method for incorporating digital twins is covariate adjustment, which Walsh’s company, Unlearn, has used extensively with clinical trial sponsors. In trials measuring cognitive test score changes, the digital twin predicts what the endpoint value would be for participants not receiving the drug. This prediction serves as a covariate in statistical adjustment.

Covariate adjustment incorporates pre-known participant data to reduce treatment effect estimate variance, providing more precise drug efficacy measurements. The digital twin’s prediction represents expected standard-of-care outcomes, not drug treatment outcomes. Walsh described forecast control outcomes from digital twins as “super covariates” providing maximum analytical power increases. Both FDA and the European Medicines Agency support using digital twins for smaller, faster, more powerful RCTs.

Walsh presented a retrospective analysis of a phase 2 Alzheimer’s drug trial with 453 participants with early Alzheimer’s disease. The model trained on more than 8,500 participants from control arms of nearly 30 past clinical trials and observational studies covering mild cognitive impairment to severe Alzheimer’s disease. The model incorporated approximately 60 variables including demographics, biomarkers, and baseline severity measures from cognitive and functional assessments.

Walsh continued, stating that results showed dramatic variance reductions compared to standard covariate analyses—10–15 percent increases in treatment effect estimate precision. While seemingly modest, digital twins explained up to 25 percent of total variance, representing significant improvement. This translates to either boosted study power with constant sample sizes or equivalent power with smaller samples.

He noted that the practical implications may be substantial. A standard phase 3 trial requiring 1,600 participants, 21 months enrollment, and more than $500 million could be reduced to 1,433 participants using digital twin variance reduction, saving 2 months enrollment time and $59 million total cost. Alternatively, maintaining participant numbers could significantly increase study power. This technology represents a paradigm shift toward more efficient clinical trial design, reducing costs and timeline while maintaining scientific rigor and regulatory acceptance.

DISCUSSION

Costes opened by noting that speakers had different definitions of digital twins; he favored Deng’s version, which emphasized an interactive system with time dependence and critical interactive aspects. He then relayed a question from the audience about creating a generic digital twin for humans that could be built upon and individualized for various purposes, such as digital cohorts.

Deng responded that this should be possible, starting with a digital twin containing different organ pieces—heart, lungs, liver, and brain—each with minimum required data capable of running basic simulations. These organs could be combined to create a generic human digital twin and then calibrated using individual patient data to create personalized versions.

Taylor agreed that creating general representations is aspirational but acknowledged the complexity. He noted that most digital twins focus on individual organ systems, suggesting that coupling various organ digital twins would be a good starting point. However, he stressed building clinically useful systems rather than pursuing perfect models, warning against spending time on frameworks without clinical application. He estimated that a full human digital twin could be a 100-year project, suggesting focused efforts to achieve wins and learn system improvements.

Deng disagreed with the timeline, suggesting that 100 years was too long for a complete human digital twin. He emphasized the value of collaboration, with different groups building individual systems and finding

integration methods. Agasthya agreed with Deng and stressed the importance of frameworks and standards enabling different groups to combine their work effectively.

Walsh viewed this as a core machine learning problem involving disparate data sources, contextualization, and predictions. He saw this as the most efficient pathway due to current investment in the area and progress in associative memory large language model approaches, although he agreed with Taylor about the 100-year timeline for complete systems.

Taylor argued for standardization, citing the Digital Imaging and Communications in Medicine standard’s tremendous advances in image analysis. He noted the current moment as ideal for building standardized frameworks for data sharing, which does not currently exist in digital twins. Regarding cardiac modeling, he noted that while his presentation focused on coronary vascular systems, more sophisticated whole-heart models are possible and already being developed, although not in patient-specific manners. He estimated that several decades will pass before clinicians can build digital heart models using all available multimodal data sources.

Deng emphasized the value of cross-disciplinary collaboration, describing his consortium’s diverse backgrounds, including clinicians. The consortium pursues two goals: developing a shared online platform repository for groups interested in digital twin development and determining data requirements for whole human models. He acknowledged that nobody knows the answer to data requirements, so they are gathering domain experts from academic and clinical sides to collectively understand minimum required data for human digital twins.

Responding to a question about incorporating dosimetry calculations into RPT, Agasthya acknowledged the controversy while outlining advantages and disadvantages. Advantages include potential patient benefits from slightly higher or lower doses than standard, such as giving higher doses to end treatment in 3 weeks instead of 6 weeks while achieving similar improvements. Disadvantages include potential over-optimization, causing detrimental effects to individuals. She emphasized that digital twins capable of predicting outcomes could enable prospective dose adjustments rather than retrospective responses to declining creatinine levels. Reliable predictions would be valuable for treatment planning.

When Costes asked about the likelihood of prospective dose shaping given strict radiotherapy protocols, Chung noted cultural resistance to protocol changes unless toxicity outcomes create concern. Agasthya agreed, explaining that standard doses are preferred for multiple reasons: manufacturer ease, simple logistics involving standard vials, and hesitancy to change doses without clinical trials proving effectiveness of alternative regimens. She suggested that even low-fidelity digital twins could provide valuable information to tailor clinical trials for testing different approaches rather than testing everything.

Walsh noted potential complications if digital twins became companion diagnostics; FDA could require model availability for any dosing decision, creating logistical challenges. Taylor agreed that this could limit drug adoption if digital twins became required diagnostics, making pharmaceutical companies reluctant to pursue such approaches. Chung said that in her experience, physician resistance seems to come from a lack of confidence and trust in new technology. Taylor agreed that building trust takes time and it can be lost quickly through single aberrant cases or adverse events. Walsh emphasized operationalization as vital and probably the biggest problem to solve. Despite developing mathematically optimal digital twins for clinical trials, he discovered that building trust, reducing risk, and demonstrating practical usability is a long process.

Chung asked about data availability challenges across different centers. Taylor confirmed this as an issue, noting highly variable data quality. Academic medical centers typically provide high-quality data, while resource-limited centers with less experience provide lower-quality data.

Quoting Stephan Achenbach, a leader in coronary CT angiography, Chung emphasized that enforcing image quality standards and rejecting low-quality cases was one of the field’s best practices. She stressed the importance of enforcing best practices for input data, citing “garbage in, garbage out,” and noted that building digital models involves consideration of input data quality, standardization, and validation measurement system quality.

Deng described particular difficulties collecting good multimodal data. Despite institutional review board approval for his Yale University study, he encounters significant resistance within the system for data access. He stated that tremendous amounts of collected data sit in individual silos or databases, making them difficult to access.

Chung ended the session by requesting final comments from each presenter. Walsh started by highlighting data standardization as important to unlocking digital twins’ potential; he described multiscale systems needing application programming interfaces across scales to model and understand systems, particularly in trials.

Taylor agreed but cautioned against overhyping AI. Despite being a “huge AI believer” who has witnessed its power firsthand, he noted current over-exuberance and warned against similar treatment of digital twins. He suggested careful model-building work instead of making 5- to 10-year predictions that will almost certainly be wrong and emphasized hard standardization work, clinical integration, and modesty about capabilities and limitations.

Agasthya agreed with both, adding that developing models that can guide people in practical VVUQ methods acceptable for understanding digital twin outputs is important. Deng felt that data access and multiscale modeling are the most challenging aspects, particularly given limited knowledge about relevant biological phenomena at microscopic levels. He emphasized the value of reaching out to domain experts for collaborative approximation-building, using real-world data to benchmark digital twin models for improved accuracy and eventual patient treatment improvements through model predictions.

REFERENCES