Mobility data improve forecasting of COVID-19 incidence trends using Graph Neural Networks (Extended Abstract)

Simon Witzke, Noel Danz, Katharina Baum, Bernhard Y Renard

Keywords: mobility data, trend estimation, graph neural networks, covid-19

TL;DR: We show that a heterogeneous GNN can use mobility data to improve COVID-19 trend forecasts for locations following a change in trend direction by leveraging information from locations where a similar reversal in trend direction occurs earlier.

Abstract: The COVID-19 pandemic has had a considerable global impact over the last few years. Many efforts were made to understand and estimate its development. The availability of large amounts of data, including mobility data, has led to numerous Graph Neural Networks (GNN) being proposed to leverage this data and forecast case numbers for the short-term future. However, information about trend developments, especially where trends reverse directions, is crucial in informing decisions. GNNs may be able to use information from regions where trends change first to improve predictions for locations with delays. We consider the first omicron wave in Germany at the end of 2021 and compare a heterogeneous GNN using mobility data with a model without spatial information. We observe that, for this period, mobility data significantly improve forecasts and specifically that improvements occur earlier in time. Using GNNs and mobility data enables leveraging information from counties affected earlier to improve forecasts for counties affected later. We conclude that such performance improvements could be transferred to counties with earlier change points by also including neighboring nations in the graph structure. Further, we emphasize the need for systematic contextual evaluation of GNN-based models for forecasting pandemic trends.