Online crowd surveillance for public health

Using Web search activity for pandemic forecasting has important implications for controlling the disease spreading and guiding the formulation of local policies. Prior studies have used Web search activity to track and predict the incidence of several infectious diseases and related location-specific models have been applied to respiratory illnesses such as influenza and most recently COVID-19. However, these models have been questioned for lacking the ability to make stable and accurate predictions, especially during the early stage of disease outbreaks. In this study, we propose a novel self-supervised message-passing neural network (SMPNN) framework to model both location-specific and cross-location dynamics for pandemic forecasting. Specifically, we apply an MPNN model to learn the cross-location dependencies in the data via a self-supervised learning process and then use the graph-generated features to enhance the prediction of location-specific regression models. We compared our model with state-of-the-art statistical approaches and deep learning models on two open datasets for COVID-19 tracking from England and the US. The results show that our model can achieve more accurate prediction with up to a 6.9% reduction in prediction errors and achieve lower prediction errors at the early stage of disease outbreaks.