Machine Learning-Based Geospatial Risk Modeling of Global Avian Influenza Outbreaks.

Abstract

The rapid spread of H5N1 avian influenza poses a global threat, highlighting the need for robust spatiotemporal risk assessment. In this study, we developed a global modeling framework integrating machine learning (ML) models and geospatial analysis to characterize H5N1 outbreak risk under varying environmental, ecological, and anthropogenic conditions. Confirmed H5N1 presence locations were extracted from World Animal Health Information System (WAHIS) (2012-2023), and pseudo-absence locations were generated using a target-group background (TGB) approach to account for heterogeneous surveillance effort. 5 ML algorithms, namely logistic regression (LR), support vector machines (SVMs), random forest (RF), light gradient boosting machine (LGBM), and extreme gradient boosting (XGB) were evaluated using spatial block cross-validation on data from 2012 to 2021 and an independent temporal holdout dataset from 2022 to 2023. Tree-based ensemble techniques (RF, LGBM, and XGB) achieved stronger and stable performance across both spatial and temporal validation. Seasonal Maximum Entropy (MaxEnt) models were applied to visualize broad-scale outbreak risk patterns across the annual cycle. Seasonal maps revealed higher risk during autumn and winter, intermediate risk during spring migration, and reduced suitability during summer, consistent with large-scale migratory connectivity, poultry production intensity, and seasonal environmental gradients. Predictor analysis indicated that livestock density and anthropogenic variables were the strongest correlates of outbreak occurrence in multivariate models, while wild bird abundance and climatic variables contributed heterogeneously and in a season-dependent manner.