Spatiotemporal Risk Assessment of H5 Avian Influenza in China: An Interpretable Machine Learning Approach to Uncover Multi-Scale Drivers

Avian influenza (AI), particularly the H5 subtypes, poses a significant and persistent threat globally. While the influence of environmental factors on AI seasonality is recognized, a comprehensive understanding of the hierarchical and interactive effects of multi-scale drivers in a vast and ecologically diverse country like China remains limited. We developed an interpretable machine learning framework (XGBoost with SHAP) to analyze the spatiotemporal risk of 1800 H5 AI outbreaks in mainland China from 2000 to 2023. We integrated multi-source data, including dynamic poultry density, Köppen climate classifications, Important Bird and Biodiversity Areas (IBAs), and daily meteorological variables, to identify key drivers and quantify their nonlinear and synergistic effects. The model demonstrated high predictive accuracy (5-fold cross-validation R2 = 0.776). Our analysis revealed that macro-scale ecological contexts, particularly poultry density and specific Köppen climate zones (e.g., Cwa), and strong seasonality were the most dominant drivers of AI risk. We identified significant nonlinear relationships, such as a strong inverse relationship with temperature, and a critical synergistic interaction where high temperatures substantially amplified risk in areas with high poultry density. The final predictive map identified high-risk hotspots primarily concentrated in eastern and southern China. Our findings indicate that H5 AI risk is governed by a hierarchical interplay of multi-scale environmental drivers. This interpretable modeling approach provides a valuable tool for developing targeted surveillance and early warning systems to mitigate the threat of avian influenza.