search

HopkinsIDD / COVIDScenarioPipeline

Public shared code for doing scenario forecasting and creating reports for various governmental entities.
GNU General Public License v3.0
104 stars 47 forks source link

How to initialize the SEIR model with fixed epidemic states? #364

Open vineetmp opened 4 years ago

vineetmp commented 4 years ago

How do I initialize the SEIR model with a known number of infections in each US county, say on 7/31/20 and generate forecasts for the period 8/1/20 to 8/31/20? I see that this document mentions that it is possible, but doesn't explain how. I know that I have to run simulate.py and that I have to set the params start_date and end_date as 2020-08-01 and 2020-08-31 respectively. I have also followed instructions provided here to set the other parameters. However, when it comes to seeding parameters/file, I can get the actual case data on 7/31/2020 from JHU CSSE, but how do I feed that into the model as the seed? Which parameter should I use?

Also, thanks to the authors for sharing the code!

eclee25 commented 4 years ago

Sorry for this oversight -- I hope to update the documentation over the next few days as we are planning to merge some new features into the master branch.

vineetmp commented 4 years ago

Okay, thank you!

eclee25 commented 4 years ago

@vineetmp It turns out that this feature was something that was implemented in a branch that did was not merged in for this release. We'll be removing that section of the methods in the preprint (thanks to your tip!). Sorry to disappoint.

vineetmp commented 4 years ago

@eclee25 Oh okay, thanks for the quick response. I thought you are using this same pipeline to generate forecasts for CDC. If that is the case, what parameters do you change when you rerun the model to update your forecasts weekly?

I ran all the steps mentioned in your documentation for COVID19_Minimal after editing the config etc. for all US counties, without any interventions (parameter set to None). And then ran the simulate.py script. When I quickly looked at the results for some US counties, the number of cases (I1+I2+I3) peak in May/June and then decline to near-zero in August. How do I update the parameters to get more realistic projections? Does adjusting seeding values/intervention settings help? Or is there something else?

eclee25 commented 4 years ago

The model runs that we submit to CDC are based on the same core model code but with substantial updates and changes to model deployment that are not very well documented yet. We run off the inference branch and the config changes are mostly documented here: https://github.com/HopkinsIDD/COVIDScenarioPipeline/wiki/Inference-model-configuration-setup. The full inference setup is deployed on AWS batch for runs of the full US. I would not recommend using the inference code locally unless you have few locations (e.g., ~10).

The model that is currently in master/dev is our scenario planning model only, which means that there is no fitting procedure to calibrate case and death time series to the model outputs.

vineetmp commented 4 years ago

@eclee25 Okay, thank you! I will look into the inference branch and see if that helps my use case.