Factor Analysis and Prediction of Disease Risk Based on Large Ensembles of Models: Application to Virus Yellows in Sugar Beet

Identifying disease risk factors, characterizing their effects, and forecasting disease risk across space and time are crucial tasks in human, animal, and plant epidemiology. Statistical and machine learning models have largely superseded purely descriptive analyses of data in handling these tasks. In addition, these models have demonstrated their full potential in the current era, characterized by an unprecedented abundance of data. However, applying these models to real-world, large-scale data sets raises critical questions: Which model should be used? Which explanatory variables should be selected? What data should be allocated for training and validation? The answers to these questions often have a significant impact on the analysis outcomes. One way to address some of these challenges is to analyze risk factors and predict risk by using an ensemble of models rather than relying on a single model. This approach is developed in this article and implemented in the case of virus yellows in sugar beet in France. Among the explanatory variables correlated with the severity of virus yellows, we identified winter and spring temperatures (positive correlation), spring humidity and precipitation (negative correlation), the proportion of cereal crops (positive correlation), the proportion of grasslands (negative correlation), and the distance to sugar beet seed production fields (negative correlation). Additionally, we found that predictions are generally more robust when using a spatial aggregation of models compared with relying on the best individual model. Our approach is highly versatile and can be applied to characterize and predict the spatiotemporal distributions of diverse diseases.