Hey there,
I’ve held this account for a long time and been reading most discussions here, love this forum and all. Figured out I should introduce myself and ideally start posting 
great little community.
I’m Joseph and I research computational epidemiology. I have a background in robotics/computational mathematics so very interested in this as well.
At the moment I’m a assistant professor at UNC Chapel Hill (within ACCIDDA which is within InsightNet)
I spend quite a lot of time thinking about scoring, nowcasting. Some example project I have right now:
- RespiLens.com is a website to view forecasts of the various forecast hubs in the US.
- influpaint is a DDPM (a genAI model similar to how chatGPT generates image) for influenza forecasting
- flepiMoP, an open-source flexible epidemic modeling pipeline that can simulate most compartmental models over a wide range of connected metapopulation setups. It is being superseded by flepimop2 that is not develop by me.
- SNIDDY, aimed to foster the community of infectious disease modelers in Switzerland.
Among threads I’ve explored in the past and would love to spend time discussing, there is optimal control of ressource allocation and inference of time-varying parameters such as R_0 inside model, and everything AI and model fitting.
3 Likes
Hi @jcblemai,
Thanks for converting from a lurker to a poster! Please do post. We are in a bit of a flux as to organisation for the forum but the general idea is anything centred around real time modelling. So anything you find interesting within that in whatever format.
robotics/computational mathematics
I’m not sure Iwas aware of this. Does this mean you know much about the signal processing literature or am I getting my wires crossed (wow great joke).
I like all of this but I really love the human judgement forecasting interface. I was talking a bit with @kath-sherratt the other day about some teaching plans for the summer and we were throwing around the idea of using something similar to get students to make human judgement forecasts as an excercise. That got me to thinking that pulling this interface out into its own little tool with some docs on converting for a given use case could be really neat…
I might just not have looked hard enough (its a long readme in a good way but leads me to think I am missing something) but is there a write up of the eval for this? I would be keen to know where it is good/bad. Also I see @adrianlison cropping up in their contributing some WIS!
Its nice to see other people following in the footsteps of the best ever naming scheme.
I’ve been speaking to @sambrand a bunch about this recently where his idea is to use Kalman filters and friends for the time-varying components with a MCMC/SMC wrapper for the other parameters. Certainly seems like a goer but hard to do without major model limitations. Something I am really keen to explore is modelling lots of time varying parameters evolving jointly.
optimal control of ressource allocation
This is not something I know a great deal about but I am keen to try out Jump in the julia ecosystem which I think has some of these ideas in it. The place I would be keen to see this hooked into more is scoring that optimal use of resources for forecasts (i.e.https://academic.oup.com/jrsssa/article/188/4/1299/7927906 from
@nickreich and friends).
thanks for the warm welcome!
yes I did quite a bit of signal processing (also embedded programming, control, computer vision… all fun I miss…). I think I did not manage to apply these to epi as well as I should have (but did some Kalman filtering/Optimal control)
About RespiLens: This is a great idea, there would be two ways to do that (an “hosted” teaching space respilens.com vs making a package. At epidemics10 there was a forecastle tournament where everyone was competing on the same 5 challenges and there was a ladder board at the end. Michael Johansson won, @nickreich was the only one with perfect coverage). Happy to discuss, do you want to open an issue on the GitHub.
About influpaint: there is no write-up yet appart from a draft. One of the main weakness is that I fit on epidemic generation (not on forecasts). I then evaluate models on the past 3 FluSight seasons (and submitted live to these sessions as well). If I were to redo it (and I might), I would fit directly on CRPS
I actually stopped using @adrianlison WIS (it has been useful for a very long time, thanks Adrian) in favor of scoringutils, which I’ve found to be great (in fact, I started by making a scoring package in python and decided to use scoring utils instead)
Re: Kalman Filter: I think there are quite a few application already, but you mean as a ready-tool ? Another Sam just released Cuthbert and really trying to find an excuse to try it. I know Julia ecosystem is stellar.
Re: Optimal control: this paper (Optimal control of the spatial allocation of COVID-19 vaccines: Italy as a case study) was really fun, a lot of what there is in literature is simple control but I believe that it’s ready for a wider use on large model/complex questions
Well I took age to reply
Perhaps unsurprisingly I would like a package but I feel like the correct answer is perhaps both?
This is exactly what I want easy tools to do on the fly.
Yes exactly. Also I saw a Cuthbert the other day - it looks great. There are a few nice options in Julia @sambrand has been losing his mind over them a bit recently.
Nice - I haven’t read this so will take a look
1 Like