Back to Main Agenda

Back to THEME 1- Ordinary Differential Equations session page

Session Lead: Mark Alber

IMAG Moderators: Elebeoba (Chi-Chi) May (NSF), David Miller (NCI)

Breakout Session Notes:

• Introductions (name and interest):
  • Session 1:
    • Ted Berger, USC. MSM U01 and developing implants for nerve disorders
    • Dong Song, USC, motion learning for prothsesis dev
    • Jim Sluka, Indiana Univ., sc RNA-seq
    • Paul Aijetan, data integration
    • Steven Lee, 
    • Bill Cannon / PNNL
    • Aaron Meyer, UCLA, dynamic models of receptor families
    • Feilim Mac Gabhann, JHU, PK-PD
    • Deb Hope
    • Greg Kelly, comp neurosci
    • Gary C An, UoVT, AI for agent based models
    • Ahmet, data management
    • Reinhard Laubenbacher  / UConn Health
    • Herbert Sauro, modeling signaling pathways
    • Mark Palmer - Medtronic / patient predictive model
    • Shayn Peirce-Cottler, organizing March MSM meeting
    • Jule Leonard Duke, UV
    • Elsje, Purdue
    • Emma Lejeune, BU
    • William Barnett, comp neurosci, wants to learn ML
    • John Bachman - Harvard - ODE model
    • Mohamed Sherif -  Yale
    • William Barnett and Ted Dick, CWRU
  • Session 2:
• Build on current state of the are ODE Models
  • Build on current state of the are ODE Models
  • Gary: Agent based models are a promising target for ML
  • ML can approx a compressed version of ABM, ML as a surrogate model of sorts
  • interest in constraints vs minimal required data for building models
  • Shayn: ABM is all about rules, handing some of the rule learning over to the ML is intriguing
  • Chi Chi: how is ML defined for this conversation?
  • how one defines rules is important
  • If the goal is control, ODE is an easier road
  • Emma: ML do not currently have established ways to handle stochasitc behavior and emergent behavior
  • Mark: scale is important
  • Gary: hybrid models need boundries around expected behaviors of the model components
  • you need confidence that you have consistant behaviors across the bioplausable behavior space.
  • an 'issue' with ML is that you may not know what features are being selected, some are not bio appropriate
  • you need a way to 'distrust' ML output
  • example: AIs have been known to exploit mistakes in code to solve problems (OpenAI Hide and Seek)
  • Ahmet: Digital twins should change practice
  • A pure 1-1 'digital twin' is science fiction, in reality a digital patient is a collection of future trajectories
  • Mark: real time feedback and updates to model is definitive for digital twin
  • A machine can be well characterized. an airplane engine has 30,000 sensors that can characterize a DT of the engine with 5 minutes of data. you can't do this on a human.
  • ChiChi: what are the current gaps?
  • William: the iterative, bottom up approach of MSM has parallels in ML
• Build on current state of the art models for Digital Twins
  •  
• ML-MSM integration opportunities
  •  
• Challenges ML-MSM modelers should address
  • a big one is that no one (or very few) has access to the data
  • Nathan Price is addressing this
  • before we can make a digital twin, we need a digital population
  • Many datasets are mostly homogenous, the heterogeneity is in the tail, which is a problem if the tails are driving the models
  • We need a more precise definition of what a digital twin is.
  • A physician's conception of a useful digital twin will be different than a researcher's
  • Mark: this group should define a DT in a healthcare context, as well as gaps in the definition re an industrial understanding of DT
  • Additional Challenges:
  • Need a near comprehensive data set for near complete dynamics for Neural Nets to train on
  • Draw backs of NNets/ML you don’t know what it knows - black box
  • Come up with a schema for implausibility for ML algorithms model
  • Need to direct the technology to acquire the proper data sets to enable forecasting capability of the digital twin idea
• Breakout Summary Discussion
  • One key aspect: We need a definition of DT. The industrial definition of DT needs to be recast for healthcare and biomedicine.
  • Herb: there's a case for developing an in vitro twin, a replica of some small unit of biology
  • A best practice is to say "A Digital Twin of ...X"
  • The point is that a digital twin should help us do something better, not necessarily understand something better
  • John: The distinction between a DT and a model is completeness and continuing to drive the model with data, allows you to make reasoning statements
  • Gary: use vs exception of DT are different in industrial vs health settings

SESSION 2/Summary

I.  Need to DEFINE Digital twin for health care
— Industrial definition may not be quite right
— Need a well posed, defined environment and develop for that condition (e.g., GE digital twin for engine not entire plane)
— Digital twin for microbiome, glycolytic particle, bacterium?
— Not just one phrase, needs to be “Digital twin of …”   - the “of” defines scope
— Digital twin should help us do something not just understand
— How does DT different from a model - DT specific instance/parameterization of an instantiation of a model; with causality ; driving it with the inputs that the physical system is generating
— Define expectations from DT for healthcare vs industrial application
—Question of  Link between DT and ML
— Mark: Inputs from real world are very disparate; analytical tools to work with that is challenging; ML very good for working with those

-- Separate of digital twin from reg modeling is that you’re feeding real time data into the system / Digital twin as layered conditions

II.  Define/recognize/quantify expectations for DT and what it is able to predict/generate
— Need to make a GFEffort to define limitation
— [Example given of prediction of glucose levels using physiological model - prediction of glucose levels and insulin needed]
— Mark to contribute examples of DT - glucose monitoring that adapts to the person

III. Challenges of extending DT from successful apps to other questions such as cancer therapy
— Mark - glucose monitoring is well posed problem
— Well posed mathematical representation of what is desired
— Feedback between DT and actual person is helpful - increases value of DT
— What is good enough?  How to define what we trust with DT.
——> Having good error bars, which will be different for industry.  Validation and testing.
— Ahmet:  DTs should change practice - will patients trust DTs
— Existance of closed loop systems for health; well posed problems (glucose, anesthesia)
—Controls - if you develop ODEs there are several tools for controls; need for ABM

IV. ABMs and ML  (significant discussion at start of session)

- Agent Based Model (ABM) as a surrogate model? Nature of how we characterize or bio plausible from an agent based model - could apply ML
— Different ML approaches - linking scale specific ML approaches imprnt

- Mean field approximation as a relationship between rules in ABM and ML
-  NNet ability for capturing the stochastic nature of the rule
-  Challenge with ABM is the PDE challenge
- Issue of scale

- Decomposing ABM to components
-  Use of ML for global sensitivity analysis for ABMs

V.  Challenges for Digital twining  (Philosophical and Technical)
-- Data access - no one has access to the data except for a few organizations / need common data formats
-- Creating a digital twin - need a digital population first
—Fundamental problem of assuming heterogeneity b/c a lot of the disease occurs at the tail of the population
-- More precise definition of what a digital twin is.
— DT is/should not be a set of phenotypes extracted …  There has to be some mechanistic info/characteristics; need underlying generative patterns
-- Industrial vs health care definition of Digital twin
— Can’t make measurements on humans similar to industrial systems/machines
-- Digital twins are we talking about humans or other things - such as microbes / digital twins for different biological processes
-- No rules for bio/cellular systems as there are for engineered systems that follow physical rules
—If moving top down maybe not possible  /Can we define the rules from the bottom up and generate the molecular rules that can be combined to generate models of biological systems
-- Need well posed questions that will help define digital twin problem
-- Terms matter