Using Neural Networks to Model the Spread of COVID-19

Abstract

The spread of the novel coronavirus across the world in 2020 exposed the tenuous nature of hospital capacity and medical resource supply lines. Being able to anticipate surge events days before they hit an area would allow healthcare workers to pivot and prepare, critically expanding capacity and adjusting to resource loads. This work aims to enable advanced healthcare planning by providing adaptive forecasts into short range COVID-19 outbreaks and surge events. Here, we present a novel method to predict the spread of COVID-19 by using creative neural network architectures, especially convolutional and LSTM layers. Our goal was to create a generalizable method or model to predict disease spread on a county-level granularity. Importantly, we found that by using an adaptive neural network model with a frequent refresh rate, we were able to outperform simple feed forward estimation methods to predict county level new case counts on a daily basis. We also show the capabilities of neural network architectures by comparing performance on different sizes of training data and geographic inputs. Our results indicate that neural networks are well suited to dynamically modeling the spread of COVID-19 on a county-level basis, but that cultural and/or geographic differences in regions prevent the portability of fully-trained models.