Towards responsible AI for plague foci surveillance and outbreak prediction: Mongolian case

Abstract Plague remains a persistent zoonotic threat in Mongolia, where Yersinia pestis (Y. pestis) bacteria responsible for plague circulate in natural foci linked to marmot-flea ecosystems. This study leverages primary field data collected from high-risk plague zones in Khuvsgul province, Mongolia, during an expedition in August 2025, combined with interpretable machine learning to improve risk assessment and outbreak prediction. Key field variables include burrow density, marmot counts, and ectoparasite loads. Due to the small scale of the curated field samples (N=54), a synthetic dataset (N=120) simulating ecological and climatic factors across a few distinct soums was integrated to enhance model prototyping to assess the natural plague foci associated with marmot populations. The augmented XGBoost model is established as the most reliable early-warning tool, offering the precise diagnostic equilibrium required to optimize targeted rural plague surveillance and field interventions. Moreover, SHAP analysis successfully recovered dominant risk drivers such as temperature and precipitation. Because these metrics are derived from synthetic data generated according to predefined ecological assumptions, this framework is presented as a methodological proof-of-concept demonstration of an explainable AI workflow rather than field-validated outbreak prediction performance. This framework demonstrates responsible AI's potential to augment resource-limited field efforts, promoting equitable public health in Mongolia and beyond.