TREA-Net Improves Dengue Forecasting in Data-Scarce Regions

This research introduces TREA-Net, a Transferable Residual Epidemiological Adaptation Network designed to improve multi-week dengue incidence forecasting, particularly in regions with newly established surveillance systems that lack extensive historical data. Traditional neural forecasting models struggle in such "data-scarce" environments. TREA-Net addresses this by augmenting existing neural forecasting backbones with projections from an Environmental Time-Series Susceptible-Infected-Recovered (ETS-SIR) model. The core innovation lies in its ability to learn a lightweight, gated residual correction that can be transferred from data-rich regions to data-scarce ones. Its node-invariant design allows it to accommodate varying numbers of locations, and target adaptation requires learning only two global parameters. Empirical evaluations demonstrated that TREA-Net significantly improves forecasting accuracy across various neural backbones and transfer settings, achieving the lowest mean absolute error when integrated with the TiRex foundation model. It also reduces prediction interval width, making it a powerful early-warning system for public health.