Machine learning maps avian influenza risk hotspots across India

A new machine-learning study has mapped areas of India where highly pathogenic avian influenza is most likely to be reported, identifying eastern, northeastern, coastal and selected southern regions as priorities for surveillance. The model combines more than 90 environmental, livestock and human-related variables and, importantly, also maps where its own predictions are uncertain.

The study used confirmed avian influenza outbreak records from January 2006 to April 2024, dominated by H5N1 and H5N8. India is particularly suited to this type of analysis because intensive poultry production, backyard flocks, dense human populations, wetlands and major migratory bird routes frequently overlap.

More than 100,000 grid cells analysed

Researchers divided India into 115,781 hexagonal cells of about 5.16 km² each. They then integrated 94 predictors, including chicken and duck density, human population density, rainfall, humidity, surface temperature, vegetation, wetlands, agricultural land and distance from permanent water.

The outbreak database initially contained 458 records, including 393 H5N1 outbreaks and 55 H5N8 outbreaks. After removing records without sufficiently precise coordinates and adding information from published studies, the final model used 71 unique outbreak locations.

Eastern and coastal India emerge as higher-risk areas

Eight different machine-learning approaches were tested. All performed better than chance, with an area under the ROC curve above 0.76. The best-performing individual model was Random Forest, with an AUC of 0.851, while a stacked Gaussian-process model reached 0.853. The difference was small and not statistically significant.

The resulting maps identified parts of eastern and northeastern India, coastal areas and several southern regions as having higher relative suitability for reported avian influenza outbreaks. Among the factors most strongly associated with model predictions were human population density, extensive chicken density, June rainfall and seasonal humidity.

Mapping uncertainty as well as risk

One of the study’s more useful features is that it does not present the risk map as a simple prediction of where an outbreak will occur. Because the analysis relied on reported outbreaks and sampled background locations rather than confirmed disease-free areas, the authors stress that the model produces relative risk or surveillance-priority scores, not absolute probabilities of infection.

The Gaussian-process model also generated an uncertainty layer. Areas with both high predicted risk and high uncertainty could therefore be prioritised for field surveillance, allowing future data to improve the model itself.

A tool for targeted surveillance rather than outbreak prediction

The study highlights an important potential use of artificial intelligence in animal-health surveillance. India cannot monitor every poultry flock, wetland or migratory bird interface with the same intensity. Spatial modelling can help veterinary authorities direct sampling, biosecurity resources and outbreak preparedness toward areas where environmental and production conditions most resemble those associated with previous outbreaks.

The authors nevertheless caution that passive surveillance can introduce bias: places with better reporting may appear at greater risk simply because outbreaks are more likely to be detected there. Correlations between environmental variables also make it difficult to interpret individual predictors as direct causes.

The value of the model therefore lies less in predicting the next outbreak than in identifying where surveillance may be most informative, while explicitly showing where the available data remain weak.