Machine learning forecasts bluetongue outbreaks up to six months ahead

A machine-learning framework that combines animal-disease surveillance records with gridded climate data identified locations and months at elevated risk of bluetongue outbreaks up to six months in advance. The logit-weighted ensemble, or LWE, delivered the strongest and most stable performance in recall-focused testing across France, Greece and Italy, although results varied among countries and regions.

The study focused on bluetongue virus, an arbovirus of ruminants transmitted by biting midges of the genus Culicoides. Researchers developed the LWE to predict outbreak occurrence at the scale of administrative units and calendar months. The forecasts used routinely available outbreak notifications and climate data, making the approach relevant to surveillance systems that already collect information on disease occurrence and environmental conditions.

Researchers evaluated the framework using bluetongue outbreak surveillance data collected from 2005 through 2024 in France, Greece and Italy. The three countries were selected for their long-running, high-quality outbreak surveillance records. The LWE was tested at forecast horizons extending from one to six months ahead, with each prediction made for an administrative unit and month. Evaluation prioritized recall, meaning the analysis focused on correctly identifying outbreak months.

Across all three countries, the LWE outperformed or matched 14 benchmark models. Its advantages became more pronounced at lead times of three months onward, when predictions were more uncertain and outbreaks were relatively rare. Predictive performance was highest in Greece, strong in France, and lower and more variable in Italy. The researchers linked these differences to variations in the consistency and geographic spread of outbreaks across the three settings, highlighting that the countries do not form a uniform forecasting environment.

Overall, the results support horizon-aware, climate-informed forecasting as a way to identify months and locations at elevated risk of bluetongue occurrence using routinely collected data. The findings indicate that the ensemble can provide stable multi-horizon predictions across heterogeneous European settings, while its uneven performance shows that geographic patterns remain important. The study evaluates forecasting performance for bluetongue outbreaks in animal surveillance data; it does not establish clinical benefits in people, nor does it show that every predicted outbreak will occur.

Subject of Research: Machine-learning forecasting of bluetongue outbreaks from surveillance and climate data

Article Title: Spatio-temporal machine learning for multi-horizon prediction of bluetongue outbreaks

Article References: Devlin, L. M., Nguyen, P. H., Cuthbert, R. N., Doan, P. N., Tran, V. H., Zhang, Z., Murchie, A. K., Bamford, C. G. G., Dick, J. T. A., Morgan, E. R., & Mai, T. S. (2026). Spatio-temporal machine learning for multi-horizon prediction of bluetongue outbreaks. PLOS Computational Biology, 22(10), e1014870. https://doi.org/10.1371/journal.pcbi.1014870

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014870

Keywords: bluetongue, outbreak forecasting, machine learning, animal health, arbovirus, climate data, vector-borne disease, disease surveillance, early warning, spatio-temporal modelling, Spatio-temporal, machine