[Discuss] Stochastic and agent-based models

[Inglezos]

Summary of this new feature

I have read about stochastic models (Imperial college covid response team uses them a lot) as well as agent-based models (https://www.nature.com/articles/s41467-020-19393-6) in epidemiology. What is your opinion on these? Could something like that be added or used to the current model implementation? Would they be better/more accurate models?

[lisphilar]

Agent-based models will be required to understand the parameter values, rho and so on. We "simulate" the number of cases in the current version, but, with agent-based model, we can "simulate" the outbreak to evaluate the effect of measures at person level.

I will read this paper after releasing verion 2.13.

[Inglezos]

So in that way the outbreak modelling would be better? And agent-based models are different than stochastic models?

[lisphilar]

FYI: https://github.com/BDI-pathogens/OpenABM-Covid19/blob/master/documentation/covid19.md

[Inglezos]

I just found this and thought to share, the page 25 of the pdf has many important and useful tools that could be utilized along with CovsirPhy or at least get ideas! https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/Imperial-College-COVID-19-Response-Team-2020-2021-Report-(22-03-2021).pdf

[lisphilar]

Thank you for your information. I checked some tools by Imperial College London.

1. , CIVID-19 UK https://imperialcollegelondon.github.io/covid19local/#map This seems to be parallel in purpose with us. Based on the relationship of the number of daily new cases and interventions, forecast reproduction number (Rt) for the next four days. We can do the same with CovsirPhy, but runtime is too long and visualization of Rt on map is not implemented.

2. CovidSIM http://covidsim.eu/ Interactively show simulated results and can change parameter values with control factors.

3. sircovid https://mrc-ide.github.io/sircovid/ More detaild SIR-derived models. Hints for find control factors.

[lisphilar]

Covasim is a Python tool for agent-based simulator. https://github.com/InstituteforDiseaseModeling/covasim

[lisphilar]

One feature could be to perform agent-based (not ODE) simulation with estimated ODE parameter values and show gif animation.

@Inglezos, Do you have any updates for this issue?

[Inglezos]

Unfortunately I did not find any new material for this. However, one interesting factor would be a country-specific analysis, extract important features and apply these to other countries for either modelling or prediction. It would be essentially a country interconnection analysis, since people move from one to another and these is virus transmissibility in such a way. This would concern the spatial distribution as well (a discussion from an old issue). And yes all this would be more appropriate to be done by agents and stochastic approach perhaps.

[lisphilar]

Thank you and yes agent approach + spatial distribution #437 will lead country/province interconnection analysis. Outbreak in one country/province may cause outbreak in another country/province because of people movement. (In Japan, outbreak in Tokyo metropolitan area occurs before outbreaks in regional citis.)

extract important features and apply these to other countries for either modelling or prediction

How can we connect this with agent approach + spatial distribution?

[Inglezos]

Hmm I am not sure. Perhaps if we try a correlation analysis for the cases or deaths between countries or even better trends-"waves". In that way hopefully, we can detect when a specific wave of x country in the past affected y country and with what delay and at what extent. Projecting this into the future could give a prediction for example the period it would hit; now if we have a wave in z country, we could expect it to propagate in this determined specific way to other countries as well and adjust the predicted parameters or cases properly. But since now we seek for patterns and features extraction, it would be a good idea to introduce neural networks for that purpose.

[lisphilar]

This may not be the same idea with you, but pattern analysis/prediction of S-R trend (or something like that) can be a new idea of analysis. Especially at province level, waves are similar each other. https://gist.github.com/lisphilar/fd04159e39b7cf4898e5509a21e31c06 (Kanagawa is a part of the Greater Tokyo Area and next to Tokyo.)