From graph models to intelligent decision-making: a review of spatio-temporal graph neural networks for regional disease risk prediction and etiology mining.

Background: Regional disease risk prediction is a core component of public health early warning systems. Traditional statistical models and machine learning methods have inherent limitations in handling multi-source heterogeneous data fusion, complex spatio-temporal dependency modeling, and interpretable etiology mining, making it difficult to meet the demands of precise and real-time public health decision-making. Objective: This review systematically examines the methodological advances, application scenarios, and future directions of spatio-temporal graph neural networks (ST-GNNs) and multi-source data fusion techniques in regional disease risk prediction and etiology mining, aiming to provide a bridging reference that connects cutting-edge technologies with practical applications for public health researchers, policymakers, and data scientists. Methods: Following the PRISMA framework, we systematically searched the Web of Science, PubMed, and IEEE Xplore databases for the period 2023-2026, ultimately including 76 core studies. A four-layer methodological framework encompassing graph construction, fusion strategies, spatio-temporal modeling, and interpretable etiology mining was developed. Results: Representative works are reviewed from two dimensions: prediction tasks (single-disease prediction, multi-disease collaborative forecasting, long-term extrapolation) and etiology mining (spatial transmission tracing, temporal pattern attribution, multi-factor interaction analysis). Five major technical challenges are identified: dynamic graph structure modeling, cross-modal heterogeneous fusion, trade-off between prediction and interpretability, out-of-distribution generalization, and privacy-preserving federated learning. These are complemented by implementation constraints from public health practice, including data availability, computational efficiency, and policy coordination. Conclusion: The main contribution of this review is the construction of a unified methodological framework integrating prediction and etiology mining, and a systematic synthesis of the challenges and future pathways at the frontier of technology and practical implementation. ST-GNNs demonstrate significant advantages in improving prediction accuracy and interpretability. Future developments should deeply integrate foundation models and causal inference to build a "prediction-intervention-evaluation" closed-loop system, providing actionable methodological references for building regional disease early warning systems and formulating public health intervention strategies globally, especially in low- and middle-income regions, thereby enhancing public health emergency response capacity and health equity.