Predicting Pathogen Emergence Using a Spatiotemporal Graph Neural Network and Explainable AI Approach

<title>Abstract</title> <p> Spatiotemporal dynamics of emerging infectious diseases continue to present significant challenges for traditional epidemiological forecasting. Driven by climate change and global mobility, outbreak patterns require advanced, transparent modeling to guide proactive public health interventions. This study presents a framework for integrating Spatiotemporal Graph Neural Networks (STGNN) and Explainable Artificial Intelligence (xAI) to forecast pathogen emergence risk. Utilising a Human Epidemic Database (1,044 events across distinct territories from 2015 to 2020), we construct a robust graph where geographic regions act as nodes connected by spatial and travel edges. The STGNN architecture captures complex transmission pathways, while subsequent xAI techniques, including SHAP, uncover the key drivers behind these predictions, such as historical outbreak counts and pathogen diversity. Across rolling temporal test sets, the STGNN model outperformed all baseline architectures, achieving an R ² score of 0.6959 and an MAE of 0.4694. Uncertainty decomposition on validation instances indicated that aleatoric uncertainty accounted for 98.85% of total predictive variance, while epistemic uncertainty contributed 1.15%. The results demonstrate improved forecasting accuracy while explicitly quantifying the relative contributions of observational noise and model uncertainty. The proposed framework will empower global health authorities to transition from passive surveillance to proactive, data driven resource allocation. Lastly, the framework presents an opportunity to help strengthen health security against future pathogenic threats, laying a vital foundation for future integration with broader climate data variables. </p>