Contact tracing algorithm

[zzj0402]

• Linear model
• Exponential model
• Pathogen model

[zzj0402]

https://discordapp.com/channels/716052909271285803/716458541837320315/717947130366394412

[zzj0402]

• https://developers.google.com/optimization/routing/vrp
• agent based SIRs with spatial movement
• potential inverse problem
• Safe Blues

[zzj0402]

potential inverse problem

[zzj0402]

Safe Blues

[zzj0402]

Zijing Zhang a.k.a. ZingToday at 5:05 PM
The problem is to give an estimation of chance being infected by the disease given a map coordinates, and time frame. For instance, I may want an estimation of me being infected given the condition that Hamilton Hospital having a CoVid 19 case(https://goo.gl/maps/ur9cjteJBdncncju9) while I am at, say Auckland airport, New Zealand within 48 hours window. I am thinking like Google Maps navigation but for virus transmission ( imagine we being the virus, what's the chance of getting a person, given conditions like washing hands/isolation rules within a time frame). I guess it's like playing Plague inc but on a smaller scale.
In short, a navigation app for the virus to infect a given person.
ChrisRackauckas from SciMLToday at 5:08 PM
I don't know if there's sufficient data to do that
Zijing Zhang a.k.a. ZingToday at 5:08 PM
I am looking at Pathogen.jl atm
ChrisRackauckas from SciMLToday at 5:09 PM
We have been looking into https://safeblues.org/ as an idea for generating real data that could possibly do it
Safe Blues
Battling COVID-19 | Think outside the box | Think beyond contact tracing

oh, you're looking to do it purely simulation based?
Zijing Zhang a.k.a. ZingToday at 5:10 PM
Yes. Maybe with some simulated data I can pull up an end-to-end neural net model
(case_location_coordinates, user_location_coordinates, time window)=> probability
ChrisRackauckas from SciMLToday at 5:12 PM
I guess if you use Pathogen you can make up a contact graph, simulate it, and get data from that kind of thing
Zijing Zhang a.k.a. ZingToday at 5:15 PM
Yes. The simulation might be primitive but I guess we can make more a fruitful one with better math after the kick-start hackathon.
ChrisRackauckas from SciMLToday at 5:16 PM
I think you might want to look at it as a potential inverse problem that could tie into Safe Blues
that's a coalition with a bunch of Aussies to essentially do a fake disease spread over bluetooth that pings back some information about when infections occur
one question you can ask is, what is the sufficient information you would need to estimate the probability of infection at different areas?
Zijing Zhang a.k.a. ZingToday at 5:17 PM
Wow, brilliant idea
ChrisRackauckas from SciMLToday at 5:17 PM
It's like the question of sufficient statistics in mathematical statistics, but applied to nonlinear estimation
and you can utilize simulations to investigate how much information you would need to solve that problem
Zijing Zhang a.k.a. ZingToday at 5:18 PM
Like a large scale Lazer tag game XD
ChrisRackauckas from SciMLToday at 5:18 PM
For example, you can create a simulation that has a real virus and then 300 of these fake strands (since some fake strands can die off and such, see the examples)
now do one of these spatial models
assume only 10% of people have the app installed: is that enough to be able to estimate what spatial clusters are "bad"?
Zijing Zhang a.k.a. ZingToday at 5:20 PM
How do we model the traffic/policy/personal hygen practice?
ChrisRackauckas from SciMLToday at 5:20 PM
start simple
start with just what you need, and then add details
the interesting question you want to address is a model with spatial locations, so the simplest model would be one of the agent based SIRs with spatial movement
now if you throw that on a google map, could you make it so people only had to follow roads, and had a rule that they had to go home?
hygiene could be just as simple as differences in transmission rates
but I'd start small and start by getting the spatial structure you want
Zijing Zhang a.k.a. ZingToday at 5:23 PM
Yes, I am thinking about borrowing some google routing algorithm add a simple virus factor as a starter
ChrisRackauckas from SciMLToday at 5:23 PM
that would be fun
Zijing Zhang a.k.a. ZingToday at 5:25 PM
I wish Facebook/social media platform has some kinda interaction score. Then I can make a simple chain of probability. Guess I can compose one just for the hackathon.
ChrisRackauckas from SciMLToday at 5:26 PM
you might be able to make one by mining Twitter
it's not very correct because, I mean, Twitter matches everyone
but :person_shrugging:
Zijing Zhang a.k.a. ZingToday at 5:27 PM
And I want to drag SciML into the equation. How do I do that?
ChrisRackauckas from SciMLToday at 5:27 PM
well without data it's hard
but there are two directions
one thing you can do is train neural surrogates to approximate the solution of agent-based models without having to run the model, allowing for real-time predictions
another thing you can do is ask questions about inverse problems like, assume you had data and a good model, could you back out hygiene practices?
Zijing Zhang a.k.a. ZingToday at 5:28 PM
Shoot, I forget about the online prediction constraint
ChrisRackauckas from SciMLToday at 5:29 PM
In general though, I think that's the wrong direction for coming up with a model
it's much easier to come up with a research question and then build a model to address it

[zzj0402]

https://developers.google.com/optimization/routing/vrp

[zzj0402]

https://developers.google.com/maps/documentation/directions/start

[zzj0402]

MCMC

https://www.youtube.com/watch?v=12eZWG0Z5gY