Machine learning can forecast bluetongue outbreaks up to six months ahead

A new machine-learning study suggests that bluetongue outbreaks can be anticipated several months in advance by combining routine surveillance data with climate information. Tested across France, Greece and Italy, the model identified periods and locations at elevated outbreak risk up to six months ahead, with the strongest performance in Greece and France and more variable results in Italy.

The study analysed bluetongue surveillance data from 2005 to 2024 and linked them with gridded climate variables. The researchers developed a logit-weighted ensemble (LWE) designed to predict whether an outbreak would be reported in a given administrative unit and month, from one to six months in advance.

A six-month early-warning horizon

The LWE outperformed or matched 14 benchmark models, particularly at longer lead times from month 3 onward, when outbreak prediction becomes more difficult and positive events are relatively rare. Performance declined as the forecasting horizon increased, but the ensemble remained comparatively stable. The authors therefore argue that the model’s main advantage is not simply peak accuracy, but its robustness across different lead times and epidemiological settings.

At five to six months, forecasts should not be interpreted as precise predictions that an outbreak will occur. Instead, they are best viewed as early indicators of increased risk, useful for planning surveillance, vaccination logistics, laboratory capacity and field preparedness.

Climate and recent outbreak history provide most of the signal

Bluetongue is particularly suited to this type of modelling because transmission by Culicoides midges is strongly influenced by temperature, seasonality and other environmental conditions. The model combined recent outbreak history with climate summaries, allowing it to capture both ongoing transmission and the broader environmental conditions favourable to virus circulation.

Predictability differed substantially among countries. Greece showed the highest performance, France also performed strongly, while Italy was more heterogeneous, probably reflecting differences in outbreak distribution and reporting patterns.

Useful, but not a complete epidemiological model

The approach has important limitations. It does not directly incorporate vaccination coverage, animal movements or detailed Culicoides population data, all of which can strongly affect bluetongue transmission. The forecasts therefore identify statistical risk rather than reproduce the complete biology of an outbreak. Even so, the study shows that routinely collected surveillance and climate data contain enough predictive information to provide meaningful warning several months before outbreaks are detected.

The potential application is to move from purely reactive bluetongue surveillance toward risk-based anticipation, concentrating monitoring and preparedness in regions where climatic and epidemiological conditions indicate increasing risk.