# Dynamic spatiotemporal graph attention networks for cross-regional multi-disease forecasting and intervention optimization

## TL;DR

This paper introduces a new AI framework for predicting disease spread and optimizing interventions by combining mobility data and machine learning.

## Contribution

A novel spatiotemporal graph attention network (ST-GAT) that integrates diverse data sources for multi-disease forecasting and intervention optimization.

## Key findings

- ST-GAT outperforms existing models like ARIMAX and LSTM in forecasting accuracy for diseases like ILI, HFMD, dengue, and RSV.
- The model identifies key transmission corridors and optimal intervention strategies that balance cost, fairness, and feasibility.
- A vaccine-first strategy is shown to be most cost-effective and stable for public health interventions.

## Abstract

Accurately predicting cross-regional spread of infectious diseases and designing cost-effective interventions is challenging due to population mobility, multi-pathogen circulation, and spatiotemporal heterogeneity. This study aims to build a unified framework that improves multi-disease forecasting, enhances interpretability of transmission pathways, and enables data-driven optimization of public-health interventions.

We develop a spatiotemporal graph attention network (ST-GAT) that integrates surveillance, meteorological, healthcare, and NPI data on a dynamic multi-relational graph combining geographic adjacency and origin-destination mobility. Spatial and temporal attention with a distribution-aware NB/ZINB decoder generates calibrated 1–4-week probabilistic forecasts, and the model is embedded in a multi-objective optimization engine to evaluate vaccine allocation and mobility restriction strategies under cost, fairness, and feasibility constraints.

Using ILI, HFMD, dengue, and RSV data, ST-GAT reduces MAE (34% vs ARIMAX, 27% vs Prophet, 15% vs LSTM/GRU) and improves WIS/CRPS across diseases. Spatial attention identifies high-weight transmission corridors, temporal attention highlights short lags of 1–4 weeks, and optimization shows a vaccine-first strategy achieves the best cost-effectiveness and stability.

The framework provides an integrated, interpretable, and generalizable solution for real-time epidemic prediction and equitable public-health decision-making.