EpiAttenNet: Dual-Stream Cross-Attention Neural Network for Multi-Horizon Epidemic Time-Series Forecasting with Uncertainty Quantification

Abstract

The problem of reliable prediction of the spread of infectious diseases has been one of the most important but unsolved issues of infectious disease dynamics, particularly when outbreaks are evolving in irregular fashion at time scales comparable to those of the epidemic process, as is the case with most epidemics. The forecasting of the spread of infectious diseases has been one of the most important but unsolved issues of the infectious disease dynamics, particularly when the external conditions are heterogeneously varying and the spreading process non-linear, which is typical for an outbreak evolving at a time scale comparable to the time scale of the epidemic process, as is typical for most epidemics. The current methods are based either on epidemiological assumptions that are only suitable for a specific compartment, which are very poor generalisers over a broad range of outbreak scenarios, or on sequence models that are very poor generalisers over a broad range of outbreak scenarios. A new Dual-Stream Cross-Attention (DSCA) architecture is introduced for simultaneously modeling epidemic case trajectories and exogenous covariates (mobility index, vaccination rates, non-pharmaceutical interventions (NPI) severity scores) that are heterogeneous to each other.For this purpose, a novel Dual-Stream Cross-Attention (DSCA) architecture is proposed, where the original temporal factors are separately encoded in parallel temporal convolutional streams, and then fused together by a learnable cross-attention mechanism. Multi-quantile regression head produces both point forecasts and calibrated prediction intervals of length 90%, thus providing the decision support regarding uncertainty information. The mean absolute error (MAE) and RMSE of EpiAttenNet are 3.12 and 4.78 respectively, representing a 24.3% and 25.8% improvement compared to the best of baseline models on time-series benchmarks with a 1-day horizon, 7-day horizon, and 14-day horizon respectively. The findings through ablation research verify each architectural member's work acting individually towards performance. The code and the pre-trained weights are released. Keywords: Epidemic forecasting, cross-attention, time-series, uncertainty quantification, COVID-19, exogenous covariates.