Predicting epidemics with mean growth rates

Three weeks after the launch of Epidemic Datathon the team "stayhome" has the highest prediction score due to their accurate case number predictions for a large number of countries. In this blog post, we want to give the members of the team "stayhome" a chance to introduce themselves and briefly describe their methodology.

About the team "stayhome"

We are Stefan Strub (25) and Hannes Löbner (24), two students who just started their PhD in the beginning of the year in signal processing of gravitational wave recordings and medical physics at DERDW and DITET at ETH Zurich. We both started studying physics together at ETH back in 2014 and except for an exchange semester abroad in Taiwan and Stockholm, we have been good friends and made it through the Bachelor and Master together. Our first programming experience was in the second semester with “Numerische Methoden für Physiker”, which opened a new and in the beginning quite difficult world to us. However finishing our Masterthesis in the subjects of simulation and optimization we feel quite comfortable now in the programming world.

Predicting case numbers

There are basically two code versions: The first version works in a script manner and is our main code, and the other one looks nice and is object oriented. Our predictions are based on the Johns Hopkins University datasets.

All codes are publicly accessible on:

https://github.com/stefanstrub/covid-19-predictions.git

Due to the current lockdown, we were not able to meet in person, and so basically everyone started on his own to get an overview of the data, how it is organized and how to handle it. The class structure of the second version is not yet implemented in the first code, but will be in the next days. As the quality of the data itself is bad, in regard of representing the actual cases (due to insufficient testing or unreliable communication), we refrained from implementing an actual model with R0, R1, … etc. Further, as a the developments in each country are quite different, and would take a lot of time to implement factors such as overfilled hospitals or age-depended demographics, we do not derive the development of one country to another. Instead we fitted exponential functions to the last 5 days of the averaged new confirmed cases for each country. This enables us to predict the average new confirmed cases and therefore predict the number of total confirmed cases.

Predicting the case fatality rate

For predicting the number of deaths we use grid search, which is a tool from inverse theory, to estimate the death latency and final case fatality rate (CRF) of the disease. Knowing these two numbers we are able to time shift and shrink the averaged new confirmed cases in order to match the average new deaths. Using this method enables us to predict the deaths based on the confirmed cases knowing that these two numbers correlate. This method has the advantage that new developments of the outbreak, for example because of a look down, which is already showing effect in the numbers of confirmed cases can hopefully predict the same development for the number of deaths even before they show an effect.