Attention-Based Recurrent Multi-Channel Neural Network for Influenza Epidemic Prediction

Abstract

Influenza is a contagious respiratory disease that can lead to serious illness. Due to its serious threat to public health, accurate real-time prediction of influenza outbreaks has a great value. In this paper, a novel deep neural network architecture is employed to provide a real-time ILI% in Guangzhou, China. Because of the long-term structure property and the diversity of influenza epidemic data, long short-term memory (LSTM) network can yield accurate prediction accuracy. We design a Multi-channel LSTM network to extract fused descriptor from multiple types of input. We further improve prediction accuracy by adding attention mechanism. This structure allows us to handle the relationship between multiple inputs more appropriately. The proposed model can make full use of information in the dataset, solving the actual problem of influenza epidemic prediction in Guangzhou with pertinence. The performance evaluates by comparing with different architectures and other state-of-art methods. The experiments show that our model has the most competitive result, and can provide the effective real-time prediction.