Leakage-safe weekly risk mapping of highly pathogenic avian influenza outbreaks in domestic poultry in Japan using open outbreak, meteorological, and geospatial data.

Abstract

Highly pathogenic avian influenza (HPAI) outbreaks in domestic poultry require timely surveillance, but national-scale prediction is challenging because events are rare, spatially heterogeneous, and influenced by seasonal and environmental conditions. We developed a leakage-safe weekly risk-mapping framework for poultry HPAI outbreaks in Japan using open outbreak records, ERA5-Land reanalysis, and grid-level geographic and land-use data. The unit of analysis was a 10-km grid-week. For each target week, all 5491 national grid cells were ranked using only static, calendar, and preceding-week information, retrospectively emulating weekly spatial prioritization. Performance was evaluated using event-grid risk percentiles and top-k capture rates. The final panel comprised 1641,809 grid-week observations from 31 August 2020-18 May 2026. In the rolling fiscal-year evaluation, mean event-grid risk percentiles ranged from 0.695 to 0.713 across four predictor sets. In the event-week evaluation of 49 strict first-occurrence-like events, the baseline Extra Trees model using geography, seasonality, and previous-week temperature captured 22.4% and 42.9% of events within the top 10% and 20% of national grid cells, respectively. The external GIS and land-use Extra Trees model increased the corresponding capture rates to 38.8% and 59.2%; the top-10% difference was 16.3 %age points (paired bootstrap 95% CI, 2.0-30.6), although the exact test of discordant transitions was not significant (p = 0.0768). Performance therefore remained moderate despite improvement with GIS information. The framework should be interpreted as preliminary regional decision support rather than as a stand-alone outbreak warning system.