Abstract
The fluctuation of human infections by the Puumala orthohantavirus (PUUV) in Germany has been linked to weather and phenology parameters that drive the population growth of its host species. We quantified the annual PUUV-outbreaks at the district level by binarizing the reported infections in the period 2006-2021. With these labels we trained a model based on a support vector machine classifier for predicting local outbreaks and incidence well in advance. The feature selection for the optimal model was performed by a heuristic method and identified five monthly weather variables from the previous two years plus the beech flowering intensity of the previous year. The predictive power of the optimal model was assessed by a leave-one-out cross-validation in 16 years that led to an 82.8% accuracy for the outbreak and a 0.457 coefficient of determination for the incidence. Prediction risk maps for the entire endemic area in Germany will be annually available on a freely-accessible permanent online platform of the German Environment Agency. The model correctly identified 2022 as a year with low outbreak risk, whereas its prediction for large-scale high outbreak risk in 2023 was not confirmed.