Infectious Disease Forecasting using Multivariate Incomplete Time-series: A Hybrid Architecture with Stacked Dilated Causal Convolutions
Brandon Mossop, Quazi Abidur Rahman · 2023
The COVID-19 pandemic brought into focus the importance of accurately predicting the future spread of an infectious disease. Hybrid neural networks combining standard convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have previously been utilized for infectious disease forecasting. This study enhances previous research by introducing a novel architecture using a Stacked-dilated-causal Convolutional Neural Network and Bidirectional Long Short-Term Memory (SCNN-BiLSTM) to forecast the spread of an infectious disease by accommodating incomplete multivariate temporal sequences. Unlike standard CNN, stacked-dilated-causal convolutions provide full coverage history, where all previous elements in the time-series input window are modelled to predict the next output.COVID-19 infection data in Ontario, Canada was selected to demonstrate the effectiveness of our approach in infectious-disease forecasting. The two main contributions of this study are the following: (1) proposing Stacked-dilated-causal Convolution (SCNN) along with Bidirectional Long Short-Term Memory to create a hybrid architecture (SCNN-BiLSTM) that can model full coverage history for infectious disease forecasting; (2) developing a multivariate model that is capable of incorporating multiple incomplete input sequences. The results obtained from this study show that the proposed architecture can successfully predict the spread of infectious disease with a long forecasting horizon by utilizing multivariate data even in the absence of complete temporal sequences for all variables.