Application of machine learning with large-scale data for an effective vaccination against classical swine fever for wild boar in Japan

Abstract

Classical swine fever has been spreading across the country since its re-emergence in Japan in 2018. Gifu Prefecture has been working diligently to control the disease through the oral vaccine dissemination targeting wild boars. Although vaccines were sprayed at 14,000 locations between 2019 and 2020, vaccine ingestion by wild boars was only confirmed at 30% of the locations. Here, we predicted the vaccine ingestion rate at each point by Random Forest modeling based on vaccine dissemination data and created prediction surfaces for the probability of vaccine ingestion by wild boar using spatial interpolation techniques. Consequently, the distance from the vaccination point to the water source was the most important variable, followed by elevation, season, road density, and slope. The area under the curve, model accuracy, sensitivity, and specificity for model evaluation were 0.760, 0.678, 0.661, and 0.685, respectively. Areas with high probability of wild boar vaccination were predicted in northern, eastern, and western part of Gifu. Leave-One-Out Cross Validation results showed that Kriging approach was more accurate than the Inverse distance weighting method. We emphasize that effective vaccination strategies based on epidemiological data are essential for disease control and that our proposed tool is also applicable for other wildlife diseases.