Initial Impressions of Epidemic Forecasting

Epidemic forecasting, by its very nature, is a direct application of computer science, math, and biology.

The theoretical basis of forecasting is in optimization, machine learning, and statistics. While various forecasting techniques are established in domains like weather prediction, the theoretical challenges in forecasting remain and are amplified in epidemiology. This is because, first, it is difficult to estimate the current state of disease spread (nowcasting). Second, the biological mechanics of novel diseases may be unknown. And third, there are many dependent environmental and social factors that influence disease spread.

Specifically, trying to understand the factors of disease spread requires forecasters to find and evaluate many, potentially unconventional, data sources across the physical and social sciences. For mosquito-borne diseases, images of puddles help computer vision models detect mosquito breeding grounds. And for COVIDcast, Delphi uses data from Doctors Visits, Google Searches, and Facebook Surveys, among others. Domain knowledge in computer science, biology, and sociology has helped us identify and understand these data.

Finally, creating clear and helpful forecasts requires significant design and software engineering efforts. As I’ve seen through COVIDcast, making sure relevant information reaches as many people as quickly as possible requires a handy knowledge of communication channels, visualization tools, and software infrastructure.

From forecasting theory to web demos, various aspects of epidemic forecasting build on advances in the physical sciences and engineering in sometimes unexpected ways. I’m very excited about the new challenges and methods in this domain and how they can improve public health.

Written on November 1, 2020