Expansion of the early warning system for avian influenza in the EU to evaluate the risk of spillover from wild birds to poultry

Abstract

Highly pathogenic avian influenza (HPAI) poses a significant threat to both poultry and wild birds. To help tackle this challenge, an early warning system for HPAI in wild birds based on spatio-temporal risk mapping, the Bird Flu Radar, has previously been developed by EFSA. This work focuses on the expansion of the existing model to assess the risk of introduction and establishment of HPAI in poultry. First, a literature review was conducted to identify the risk factors for virus introduction from wild birds into poultry farms and the availability of associated data in Europe. Second, a theoretical modelling framework was developed to assess, on a grid of 50 x 50 km cells, the relative weekly probability of HPAI introduction in at least one domestic poultry flock because of infectious wild birds. This probability was estimated as the combination of two probabilities: the probability of HPAI entry into the flock and the probability of HPAI establishment in the domestic poultry population. The model outcomes are computed for all farms together but also for twelve types of farms separately. Farm types were defined based on their production type and poultry species kept. Italy and France were used a case study to test the model performance over one year of data (February 2023 to March 2024), comparing model predictions with outbreaks reported as primary outbreaks in the European Union (EU) Animal Disease Information System (ADIS). For Italy, the model performances were good, with all the outbreaks being detected in areas within or close to high-risk spatio-temporal units. The results obtained for France were more mixed: several outbreaks were reported in high-risk areas, but some were missed, apparently due to the high influence of some key model parameters and geographical specificity. Indeed, all the outbreaks reported in Southwest France were not predicted by the proposed model. These first results are encouraging, but future work should focus on finding ways to adjust certain model parameters and to improve the assessment of model performance considering a longer time period and/or including more robust input data.