Spatiotemporal Disease Case Prediction Using Contrastive Predictive Coding

Abstract

Time-series prediction models have played a vital role in guiding effective policy making and response during the COVID-19 pandemic by predicting future cases at the country, state, and county levels. However, time-series prediction has traditionally been approached as a supervised learning task requiring large datasets to learn from. The urgent nature of novel disease control means that researchers and policymakers often do not have the luxury of having robust data. Therefore, we propose Spatial Contrastive Predictive Coding (SCPC), an unsupervised learning approach to extract useful representations from high-dimensional data. In addition to time-series COVID case and death data, SCPC incorporates a covariate mobility matrix into the loss function, representing the relative number of individuals traveling between each county on a given day. The learned parameters of the SCPC model are converted into vector form and are passed as input into a multi-layer perceptron (MLP) to give a final prediction of the number of cases for each county seven days into the future. Preliminary findings show that SCPC applied to COVID-19 data achieves a low mean squared error. In future work, we plan a comparison to baseline time-series prediction models to quantitatively evaluate the competitiveness of SCPC.