THEME 3 - Data-Driven Breakout - Digital Twin

Back to Main Agenda

Back to THEME 3- Data-Driven Approaches session page

Session Lead: Bill Lytton

IMAG Moderators: Elebeoba (Chi-Chi) May (NSF), Ken Wilkins (NIDDK)

 

Breakout Session Notes:

  • Introductions (name and affiliation): interest stated verbally
    • Session 1 (with some Session 2 additions):
      • Bill Lytton - SUNY Downstate
      • William Barnett - GaState
      • Ted Dick - Case Western Reserve U
      • Chase  Cockrell - UnivVermont
      • Ashlee Ford Versypt - OSU
      • Reinhard Laubenbacher - UConn School Med / Jackson Labs
      • Jacob Barhat - Independent
      • Ahmet Erdemir - Cleveland Clinic
      • Yiling Fan - Massachusetts Institute of Technology
      • Mark Palmer - Medtronic
      • Kristofer Bouchard - 
      • Haroon Anwar -
      • Mohamed Sherif - Yale
      • Elsje Pienaer - Purdue
      • J. L-Duke - UVa
      • Amy Gryshuk – Lawrence Livermore National Laboratory
      • Amanda 
      • Susan Wright- NIDA
      • Tong Song - USC
      • Simon Giszter - Drexel 
      • Gunnar Cedersund 
      • Mark Alber - UC-Riverside
      • Andrzej Przekwas - CFD Research Corporation
      • Amir Barati Farimani - Carnegie Mellon
      • Dong Song - U. Southern California
    • General Comments
      • AFV- Point of semantics and how “Data Driven”  approaches are viewed by agencies.
        • BL - related to the connection to MSM, certainly agent based applies
        • Discussion of meaning/why data driven  
        • BL - Contrast with theory driven (vs Data Driven)
      •  Jacob - what type of data is available
      • -- TD - described data type producer 
      • --WB - as consumer of data, discussed goals
      • --BL - What percent of NIH funded projects generate public data
      • -- Discussion of usable data and formating in a way that is usable
      • -- Ken - discussion of FAIR principles for generating database repositories that are actually used by the community; want to also share models as well as data
      •   --Jacob - discussed things that deter use of public databases - manipulating data into a useable form
      • -- Ken - points to metadata issues (mentioned NLM efforts to make things interopperable) / To add a link / Discussion of standards (e.g. FHIRE)
      • --SW - NIH announcement related to FHIRE
      • --BL - Discussion of ModelDB.  Issues there related to metadata for models
      •   
      • DATA DRIVEN vs THEORY DRIVEN
      • CC - draw distinction between data driven and statistics driven models
      • BL - ML tends to be iterative
      • KW - task at hand and what you're trying to drive towards and how that determines methods used; different methods
      • ABF-- Models for molecular dynamics for protein modeling
      •   
      • LOOSE ENDS FROM THE ODE Day 1 Discussion
      • AFV - ABM simplified to ODE, but others using it for spatial scale resolution and more comparable to PDE
      • MA - Emphasize stochastic aspects of ABM
      • AFV - stochastic based methods - how ML connects more to methods
      • ABF - ML connecteded to generative vs ___ models; learning stochastic PDEs using ML
      • MA - getting access to medical data.  Huge challenge.
      • KW - not only those that hold the data but those that are contributing the data - they have a stake and responsibility to that group as well.
      • Jacob - restriction on medical data (HIPA) and messy data in clearing house sites.  
      • KW - discussed some NLM efforts to investigate data issues related to clinical trials/studies (to add link)
    • Build on current state of the art  Data-Driven Models
      • BL/KW - early state, what we need next is better databses, better integration of DBs 
      • KW - How to integrate animal model data with human model; particularly where  animal data informs 
      • MA - Digital twin for Animal
      • GC - animal is 4th M in M4 - have rodent version of model; also working on organ on a chip
      • MA - do we move from microfluidic to animal to blood?
      • GC - modular approach - sort pieces based on functional panel then use iterative methods to characterize in various species then can move to species comparison;
      • GC - feature specific comparison for well characterized features
      • GC - need to solve reproducibility not just in models but on the experimental side; using systems approach to assist with this
      •  KBouchard  -  what is the incentive to enabling cooperativity in considering reproducibility
      • KW - journal/publication based motivation
      •  ABF - re transferability of models (animal-human-patient) - (methods to address) transfer learning methods, multi-task learning methods; design architecture where the transferability features can be associated with params
      • Jacob - govt role in reproducibility
      • KW - need not just "talk" but culture shift to address reproducibility
      • GC - re the incentive portion - once systems level becomes standard for integrating data/if that becomes the expected (scientific norm) then reproducibility will solve itself
      •   KB- what do we mean by a useful ML model needs to be considered.  In general do not gain insight into structure of data but reproduce it very well.  Is that satisfactory?  Have we learned about bio
    • Build on current state of the art models for Digital Twins
      •  KW - Methods involved in digital twin
      • GC - requirements for data and use of data is quite different for mechanistic models vs ML methods;
      • GC - once a phenotypic observation in data, that is sufficient to start mechanistic model (very little data as long as your observations are reproducible)
      • GC - ML requires significantly more data b/f getting started
      •  KB - at what level of resolution are we defining digital twins
      • MA referenced and summarized discussion from ODE breakout on Day1
      • BL - application dependent - referenced hypertension
      • Jacob - question of resolution is related to how accurate do you want to be - and question to agencies is what is good enough -
      • Comment - the accuracy question has many stakeholders 
      • MA - the question of dynamics is impt - b/c digital human under some type of stress/condition/treatment
      • GC - setting the expectation in the right frame is key.  Need to avoid falsehood of inability to be 100% accurate so why try.  However current clinical practice is oftentimes pedagogics
      • KW - true for several diseases that are highly associated with lifestyle changes
      • GC - problem of adherence but use of simulation could motivate adherence
      • MA - may not be as effective with elderly.  Discussed behavior and that is a key factor
      •   - Why would patient respect the authority of digitial twin vs doctor
      • BL - frequency of digital twin updates and reminder vs less frequent interaction with doctors
      • GC - Digital twin can help motivate discussion around behavior that impacts health and less expensive and more accessible than doctor/health specialist.  Leads to participatory medicine.
      •  KB - If public not as in tune with science and questioning fact-based studies, how much will they believe the predictions
      • GC - trust will build over time 
      • KW - think of weather forecasting -- population relatively trusts
      • MSharief - how we use the digital twin
      • GC - should not dismiss becuase of not having 100% accuracy (refer to discussio from yesterday)
      • Simon - US implemntation may be more problematic (vs Sweden) and protection of information from insurance companies, etc.  Security and management of data related to DT is an issue.
      • MA - car insurance companies and use of ML and how that relates to decisions that affect individuals
    • ML-MSM integration opportunities
      •  
      • MA - digital twin for research vs clinical would be different; digital twin has been created for specific conditions; doctors/clinicians are conservative 
      • MA - gap between clinical community and comp modeling community
      • MA/KW - there are challenges to bring in DT into clinical practice (ethical and legal issues, etc.)  - need champions within clinical realm
      • KW - issue of getting access to data for underrep populations
      • BL - DT as electronic diary that informs clinicians (MA - useful for monitoring elderly)
      • GC - As we heave more personal monitoring devices, public will expect clinicians to take that into account.  DT can provide summary of that data and provide a state variable describing patient status
      • -- Why not build model that replaces reliance on clinician  (KW - expert systems)
    • Challenges ML-MSM modelers should address
      • MSharief - physician may use ML and multiscale modeling to help inform decision
      • KB-- Google satisfied with black box model; what would be great is to be able to have predictive model and be able to extract that knowledge 
      •  KB-- explainable AI is different; want to know the representation that the network is learning - intermediate representation constraining those
  • Some resources mentioned during discussion
Table sorting checkbox
Off