STANN: Spatio-Temporal Attention-based Neural Network for Epidemic Prediction

Abstract

Accurately forecasting the future trend of an epidemic plays an essential role in making effective and efficient public health policies for disease prevention and control. In reality, the local transmission of an epidemic depends not only on the cumulative number of infections at the same location, but also on the geographical spread of the disease from nearby locations. Therefore, the epidemic data usually show high nonlinearity and certain spatio-temporal patterns. Most existing methods lack the ability to simultaneously characterize the dynamic spatio-temporal patterns, thus cannot make satisfactory prediction results. In this paper, we propose a spatio-temporal attention-based neural network (STANN) to solve the epidemic prediction problem, where attention mechanisms are adopted to effectively capture the dynamic correlations of epidemic data in both spatial and temporal dimensions. The architecture of the network consists of three modules: a temporal attention module, a spatial attention module, and a temporal convolution module. Experimental results on the epidemic prediction of malaria cases in Yunnan Province, China, demonstrate that the STANN model outperforms the state-of-the-art baselines.