Deep Learning Assessment of Pathogen Transmission Across Wildlife Livestock Interfaces

Abstract

Zoonotic disease spillover at the interface between wildlife and livestock represents a serious economic problem for the farming industry and public health challenge worldwide. Classical methods of epidemiology such as conventional compartmental models always fail to account for complex non-linear environmental contact patterns and dynamic interactions between multiple species involved. In order to address this gap, this research proposes a novel approach based on the use of Spatio-Temporal Graph Neural Network (ST-GNN) with attention-based mechanisms to evaluate and forecast the risk of cross-species disease transmissions. The architecture successfully combines heterogeneous data types including high-resolution GPS telemetry, images captured by cameras traps and meteorological covariates to form a dynamic contact network wherein nodes stand for animal groups and edges represent proximity events. On the multi-species interface tracking datasets, the model shows a highly accurate performance with an 94.2% classification and 0.96 AUC-ROC. Moreover, spatial risk assessment based on the model is able to identify high-risk areas where domestic and wild animals' pathways cross each other. Empirical results reveal that spatial analysis powered by deep learning methods can change the landscape of early warning systems helping epidemiologists, veterinarians and wildlife management to predict prospective outbreaks