Machine-learning algorithms to identify key practices contributing to variation in biosecurity levels on Japanese...

Transboundary diseases, such as African swine fever, pose significant threats to the Japanese swine industry, making robust biosecurity essential. However, established assessment tools quantify overall biosecurity levels, and the specific practices driving performance variation among farms remain unclear. This study applied machine learning algorithms to identify key biosecurity measures distinguishing high-performing farms. Total, external, internal, and diagnostic monitoring scores from 254 records of the biosecurity assessment tool (2014-2023) were analyzed using Random Forest models coupled with Boruta feature selection, incorporating the assessment year to adjust for temporal shifts. The models demonstrated robust predictive accuracy (R²=0.83-0.93 for total, external, internal, and diagnostic monitoring scores). The analysis identified the rigorous management of high-risk external interfaces-specifically, visitor shower-in requirements, strict entry criteria, and the use of exclusive slaughter transport vehicles-as primary determinants of farm-to-farm variation. For internal biosecurity, controlling access to shared spaces like break rooms, alongside segregating production stages via clothing and boot exchanges, emerged as key differentiators. Notably, overall and diagnostic monitoring scores were strongly driven by the "software" of biosecurity-specifically, the strict maintenance of implementation records (e.g., for vermin control and disinfection) and adherence to sampling manuals. These findings suggest that biosecurity variation in Japan is defined by the integration of robust physical barriers with highly systematized, verifiable management protocols. To elevate the national biosecurity baseline, stakeholders should prioritize investments in external risk interfaces and the rigorous documentation of daily practices.