Modeling the spread of COVID-19 at a high spatial and temporal resolution has become an important task in the public health response to the disease. For instance, accurate county-level forecasts are not only central to monitor the state of the pandemic but are also important to efficiently allocate scarce resources such as ventilators, personal protective equipment, and ICU beds. These forecasts can also help make progress towards efficient early detection systems.
However, forecasting COVID-19 poses unique challenges – in particular when considering confirmed cases at high spatial resolution. Although there has been considerable progress towards understanding the spread of the disease, only limited data and knowledge exist about important factors that influence its spread. This is only exacerbated by the naturally larger noise-levels in county-level data as compared to more highly aggregated state-level data. Due to the global nature of COVID-19, the available data is also distributed among regions with very different properties, many of which may affect its spread. This includes, for instance, demographics and population densities, enacted policies, adherence to those policies, mobility patterns, and geographic features such as temperature. In addition, testing and reporting can change across regions and time. All these factors lead to variability and uncertainty in the data and make reliable forecasts at high spatial resolution difficult.
To improve the quality and robustness of COVID-19 forecasts, Facebook AI developed a new neural autoregressive model that aims to disentangle regional from disease-inherent aspects within these datasets. Central to this model is its ability to account for relationships between different counties; for example, an uptick in one area can have an impact on predictions for adjacent or similar districts. This spatial approach allows Facebook researchers to train models toward solutions where knowledge about the spread of the disease in one area can improve predictions in a different area and thus borrow statistical strength across counties.
In experiments, we observed that the Facebook AI method achieves strong performance in predicting the spread of COVID-19 when compared to state-of-the-art forecasts. This method takes a unique, highly data-driven approach with fewer modeling assumptions as compared to, for example, very detailed and mechanistic compartmental models. As such, we see this approach as complementary to existing models with focus on strong forecasting performance at the cost of reduced interpretability.
Learn more about our methodology by reading the full paper here.