Predicting highly pathogenic avian influenza H5N1 outbreak risk using extreme weather and bird migration data in machine learning models.

Background: Climate change is intensifying extreme weather events (EWEs) with potentially profound consequences for zoonotic disease dynamics, yet the mechanisms linking EWEs to highly pathogenic avian influenza (HPAI) H5N1 outbreaks remain poorly characterized. The ongoing H5N1 panzootic - responsible for infection in over 500 avian and mammalian species, as well as nearly 1000 human cases and 477 deaths worldwide - provides a critical opportunity to evaluate how climate conditions shape spillover risk at landscape scales. Methods: We compiled a county-month dataset of confirmed H5N1 detections across the contiguous United States from 2022-2024 and integrated it with satellite-derived climate metrics, storm event data, and wild bird activity data. We trained and validated a gradient boosting machine classifier to predict outbreak risk and characterize predictor relationships. Results: Our model achieved strong discriminative performance (AUC-ROC = 0.856; AUC-PR = 0.237, representing a 7-fold improvement over chance) and high recall (0.726), supporting its utility as an early warning tool. Human population and temperature-related variables were the most influential predictors: cold temperature shocks and prolonged low temperatures were consistently associated with elevated outbreak risk, likely through enhanced environmental viral persistence, wild bird habitat compression, and allostatic stress-driven immunosuppression in reservoir hosts. Among storm variables, high wind coverage elevated risk, potentially via aerosol dispersal of contaminated particulates, while tornado activity showed an inverse relationship, consistent with documented avoidant behavior in migratory birds. Wild bird reservoir density showed a strong positive monotonic relationship with outbreak risk. Conclusions: Our analyses demonstrate that routinely available environmental and infection data can be used to predict HPAI outbreak risk at fine spatiotemporal scales. These findings demonstrate the divergent roles of short- versus long-term environmental exposures in HPAI spillover dynamics, as well as the potential for machine learning-based surveillance tools to inform targeted biosecurity interventions and early warning systems.