Breakout 1 - Digital Twins

Back to 2023 Agenda

Back to main Breakout Session page


1. Digital Twins: whole person, mental health

MSM lead: Gary An; IMAG lead: Liz Ginexi

Proposed Working Definition of Medical Digital Twin

The term "medical digital twin" is currently used with multiple definitions. For purposes of this breakout session, we would like to provide a working definition tailored to Multiscale Modeling and uses the minimally sufficient criteria for a "digital twin" as described by Grieves in 2019 for industrial applications (recognizing that functionally this process had been in use for several years at NASA). These criteria are essentially:

  1. A data structure for the real-world object/system that would allow a specific real world object/system to be represented by a personalized/individualized virtual object or ensemble of object. (Twin structure)
  2. Representation of some process that links time points in the virtual object (Twin behavior)
  3. A link between the real and virtual twins that provides an ongoing data stream to update the virtual object (Twin Connectivity)

Operational Classification of Medical Digital Twins based on intended use:

  1. Monitoring/Diagnosis/Prediction/Forecasting
  2. Optimization of existing therapies
  3. Development/testing of novel therapies

Because this is a meeting of the "Multiscale Modeling Consortium" we would like to emphasize discussions regarding Medical Digital Twins that utilize/require mechanism-based multi-scale models

Some Pre-emptive Answers to Commonly Asked Questions:

  • Q1: What is the difference between a model and a Digital Twin?
  • A1: Most would say that the difference is that a Digital Twin includes a model (#1 and #2 from above) with a data stream from the real world twin that updates the virtual twin (#3 above). The update-ability of the Digital Twin is a key point. See Ref….
  • Q2: What is the difference between virtual populations and populations of digital twins"
  • A2: We would suggest that the difference is the personalization capability of a digital twin. While methods for generating virtual populations can theoretically encompass the heterogeneity present across a population, there is generally no explicit goal of representing a specific individual. This does not mean that methods for creating virtual populations cannot be used to create digital twins, but rather that an application may not be explicitly specified to do so.
  • Q3: What is the difference between a personalized predictive model and a digital twin?
  • A3: While this may be a potential point of disagreement, one could point again to Criteria #3 above to make the distinction: with a Digital Twin, there must be recurrent and ongoing data feedback from the real world to update and refine the future behavior of the virtual twin. For example, personalized predictive models, such as ones that utilize genomic profiles of tumors to suggest therapies specific for those tumors, while incredibly useful, only become digital twins if the tumor/patient's response can be fed back into the virtual object such that future behavior of the real world twin can be projected/forecast. We would hope that rather than being a point of contention, this distinction would be a starting point for discussion about desirable future capabilities that can turn personalized predictive models into digital twins.
  • Q4: What is the difference between virtual tissues and a medical digital twin?
  • A4: Just as an individual is made up of multiple tissues, so too should our aspirational medical digital twin be made up of multiple computational representations of its tissues. While the eventual goal would be an integrated virtual representation of a complete person, we recognize that in the path towards this aspirational goal it will be necessary to identify intermediate points where the utility of the digital twin concept (specifically the ability to utilize ongoing data links between the real and digital world) can be demonstrated, and that these intermediate points will involve the representation of specific disease processes that employ a subset of integrated virtual tissues. As with our pre-emptive answer to Q3/A3 above, rather than becoming a point of semantic contention, we hope that the perspective that digital twins represent integration of multiple virtual tissues will provide a starting point for discussions related to the modularity and composability challenges present as open research questions.


Agenda/Discussion Topics for this Breakout Session

  • How would the MSM Consortium play a role?
  • More specifically, where can the MSM Consortium fill gaps in existing large-scale initiatives regarding the development of medical digital twins (such as the European projects: the European Virtual Twin/EDITH project and the Neurotwin project
  • How to integrate mechanism-based multi-scale modeling with machine learning/artificial intelligence to develop medical digital twins?
  • How to foster collaborations that integrate underlying digital twin model development (Criteria #2) with developers of sensor/assay technologies needed to fulfill Criteria #3?
  • Propose strategies for being able to capture inter-individual heterogeneity in a way that allows personalization of digital twins.
  • Propose strategies for dealing with uncertainty and parameterization when integrating models/modules that cross multiple scales.
  • Identify strategies to find useful "intermediate points" for medical digital twin applications while still moving towards a larger vision of an integrated, whole-body medical digital twin.



Digital Twins resources




      Real World use cases justifying medical digital twins

      This comment is adjacent to the ubiquitous terminological issues we attempt to preempt with our Commonly Asked Questions above, and it pertains to real-world circumstances that industrial digital twins are used/required for, and what those scenarios might look like the biomedical domain. Specifically, in industrial applications, it is the ability to project the future trajectories of the real world twin following unanticipated events. For instance, with an digital twin of a energy plant, the representational capacity of the digital energy plant is able to deal with any conceivable perturbation to that system that affects it operations: the direct effects of natural disasters, disruption of supply chain for maintenance, terrorist attacks, etc, all of these are representable given the data specification of the digital energy plant. Another example would be unexpected changes in the operating conditions for a jet engine, such as the need to fly through a more particulate atmosphere if there is a volcanic plume that intersects its flight path, or if there happens to be an unexpected increase in bad weather that the specific airplane has to fly through. The capability of industrial digital twins to do this is predicated on the ability of the IDT to "reset" in response to a change in the data stream from the real world. One such analogous example in biomedicine is a cancer patient, already undergoing a treatment regimen proscribed by a personalized predictive model, gets into a car crash and suffers injuries that require several surgical interventions. While there exist "general" recommendations regarding the timing of chemotherapy in the face of operative interventions, what is the best strategy for this particular patient? Another case would be an manifestation of mental health issues that arise because of the stress of dealing with the cancer and its treatment: how would this affect the efficacy of a particular treatment regimen, and perhaps the best treatment for this particular patient might be an treatment strategy that accounts for the generation of mental stress? Unless that personalized predictive model (as useful as it is) could be "reset" to provide personalized prediction in the face of such an unanticipated event, it would not be able to fulfill the function (by itself) of a digital twin. Note that this is not saying a medical digital twin must be able to represent the future trajectory of every possible thing, but rather it should (aspirationally) be able to incorporate "any" (within reason...) perturbation that affects what the ostensible target of the digital twin would be, whether that is maintaining energy output (the energy plant), sufficient thrust (the airplane engine) or control of a tumor (the cancer patient). 

      The additional insight that arrives from this functional-capability criteria for a digital twin, is that it emphasizes the need to contextualize whatever specific disease-related metric is the focus of the purpose of the digital twin into the whole patient, even if the entire patient is not represented at the same degree of mechanistic granularity as the disease-specific component. Thus there needs to be an ability to represent the bi-directional effects and consequences between the disease-specific representation (e.g. a virtual tissue) and the entire patient. This relationship, and the degree of detail sufficient to represent the relationship, should be kept in mind when identifying "intermediate" use-cases during the development/deployment lifecycle of medical digital twins.

      Submitted by Gary_An on Thu, 06/22/2023 - 18:01

      Difference between model and MSM DT: Data Link

      From Reinhard L: Beause the MSM MDT is intended to have an ongoing data link to the real world twin and is expected to be able to be updated by that data stream, the design of the underlying MSM for a MSM MDT must take into account what type of data can feasibly extracted from the real world twin. While what this data is depends on what the use case for the MSM MDT is (for instance, tissue information may be more readily available in MSM MDTs for cancer) this practical constraint influences the design of the underlying MSM, and inherently introduces issues of cross-scale calibration/non-falsification that should be accounted for. Also note, given the aspirational nature of MSM MDTs, part of the purpose of MSM MDTs would be to suggest what sort of data would need to be extracted from the real world twin, and be used to provide model-driven sensor/assay development (this in turn implies that a consortium aimed at developing MSM MDTs should engage with potential developers of such assays). As with all things, there is a bidirectional relationship between data type and the MSM; development of the MSMs under the MDTs will also need to pursue the ability to have the MSMs be able to generate (in a reliable fashion) metrics that correspond to the sensors/assays.

      Submitted by Gary_An on Fri, 06/30/2023 - 08:28

      Some use-specification protocol for MSM MDTs

      This concept is an expansion of #4 from the 10 Credible Rules and analogous to the ODD Protocol for ABMs from Volker Grimm. Assertion: We have a reasonable expectation that the first generation of MSM MDTs will not be general purpose MDTs (for a host of reasons...); rather they will likely focus on some subset of disease processes. As such there is will be an inherent 'fit for purpose" component in the initial design of the MSM MDT. It would therefore be extremely useful that such there be an explicit (as much as possible) statement as to the purpose of the particular MSM MDT. In addition to a particular use-case specification, there are two other aspects that would be inherent to MSM MDTs:

      1) since the primary exclusive capability provided by MSM MDTs is their ability to evaluate/test/predict the effect of interventions that are novel to biological context (i.e. new drugs, new combination of drugs, repurposing of existing drugs), there is an inherent supposition of a control target in the design of an MSM MDT. It is therefore reasonable that the lowest scale represented by the MSM MDT must be at least at the level of that potential control. The natural example is the effect of a molecular entity (e.g. a drug): the effect of that molecular intervention will affect the behavior of cells (and, scaling up, cellular populations): in this case representation of that level of interaction should be the minimally sufficient lowest scale of an MSM (NOTE: This does not preclude finer grained representation but such representation comes with it the aforementioned issues related to calibration/non-falsification of these lower scale generative processes).

      2) since the MSM MDT is linked to the real world in terms of some metric correlating to the presence/severity of the target disease, the highest level/coarsest grained representation of the MSM MDT should be able to generate these metrics as the MSM output. This is less of an issue for the MSM community, since this cross-scale translation is inherent to the MSM task, but it would be beneficial to have this explicitly specified (as much as possible).

      3) since we recognize that the real world twin is in fact an entire person, it will be necessary to contextualize the specific disease focus of these 1st gen MSM MDTs to the entire person. We recognize that in this 1st Generation it is unreasonable and intractable to attempt to represent the entire person at the finest level of granularity (though this may be aspirational), the choices of selective abstraction should be noted as explicitly as possible, even if the initial choice to abstract those elements away.

      4) use-case defines the temporal scale of what "real time" means for the MSM MDT. There is reasonable expectation that the temporal scale of mechanism representation within the MSM MDT is at a finer grain than the data streams available from the real world twin. However, what that interval is will depend on the particular use-case in terms of relevance to the biology (and putative control) of the targeted disease. We make this statement to aid in potential misunderstanding of the phrase "real time." This "real time" is linked to the above biological/disease process: it could be as short as minutes/hours (e.g. perhaps for critical illness) or as long as months (e.g. for cancer). This updating component inherent to a MDT is determined by the sense/actuate/response-evaluation modalities for a specific disease, and is intimately tied to the required property that MSM MDTs be linked to the real world (Grieve's Def Pt 3). NOTE that exploratory MSM MDTs can be used to help define what sort of sensors/controls (and their respective turn-around times/application frequency) as part of a development consortium that would include the developers of such capabilities.

      Submitted by Gary_An on Fri, 06/30/2023 - 08:45

      ANN surrogates: online/ vs offline and "real time"

      One proposed architecture of a MSM MDT that integrates MSMs and ML/ANNs:

      There is a reasonable expectation that the MSMs required for MDTs will be necessarily complex, computationally-intensive models. Notwithstanding the aforementioned issues of disease specific "real time", it is likely that for the foreseeable future it will not be feasible to run the entire MSM for every patient/potential patient. Therefore, one potential strategy is to use the MSM to train surrogate ANNs that can be used for more of the "real-time" execution of the MSM MDT. In such a case there might be a 'fault detection" feature that defines when the ANN surrogate starts operating outside of its training set; this would then trigger, perhaps, the need to then call the the larger MSM in order to explore the new potential behavior space of that individual. Ideally, the initial training data generated by the MSM would encompass as much of the potential trajectory space of the targeted disease (and putative controlled space), but this potential transition should be accountted for (particularily with exploratory MSM MDTs).

      Submitted by Gary_An on Fri, 06/30/2023 - 09:03