Modeling the spread of highly infectious diseases, such as the COVID-19 currently menacing many nations around the world, is crucial for determining policy and informing the public. Yet the models are necessarily subject to two sources of uncertainty: first, the properties of the infectious agent itself, such as: how likely it is to be transmitted when two or more individuals congregate; how long it takes between transmission and an individual becoming themselves infectious; the rate at which infected individuals themselves are known as opposed to transmitting the disease undetected; etc.
The second uncertainty involves the movements and interactions of individuals themselves, and the way these habits are influenced by government policies. For example, in the US, even within states there is great variability as to how different cities are implementing and enforcing stay at home policies; and, so far, travel between constituencies, even though discouraged, still occurs, further complicating the modeling task. In other nations, restrictions of higher or lower severity are being implemented and also change with time as governments respond to accelerating or decelerating infection rates.
To address this bewildering complexity rapidly and yet practically for use by governments and the public, we are taking two complementary approaches. The first approach uses "bulk" mathematical models of infection, i.e. averaging across populations, coupled with a deep learning module. The role of the latter is to interpret publicly available infection rate data as outcomes of the policies and citizens' response. On its own, this would not be possible either using bulk analytical models alone or using data science alone; however, preliminary results show that coupling bulk and data-driven models can give yield quite accurate predictions of how infection rates evolve. The second approach is more granular, creating a number of representative simulated individuals and using the available data to match their behavior to real-life behaviors in a statistical sense. This approach allows us to also model the effects of transit lines, traffic, and congregation habits, etc. as well as government responses such as testing rates and quarantine policy, on epidemic spreading, to help develop better infection management practices.
[1] Neural Network aided quarantine control model estimation of COVID spread in Wuhan, China
Raj Dandekar, George Barbastathis
arXiv