Quantile regression forest-based early warning of Salmonella foodborne risk using multi-source food safety surveillance data.

Salmonella is one of the most common foodborne pathogens worldwide, with transmission closely linked to contamination across multiple food products. Effective early warning of Salmonella risk is essential for food safety management and timely intervention. However, conventional statistical models often perform poorly when surveillance data are limited and highly heterogeneous. In this study, we propose a machine learning-based early warning framework using quantile regression forests (QRF) that integrates multisource food contamination monitoring data with human health surveillance data collected from several Chinese cities between 2004 and 2023. Four tree-based models-decision tree (DT), random forest (RF), gradient boosting (GB), and QRF-were developed and compared for predicting foodborne Salmonella risk. Among these models, QRF achieved the best overall performance (MSE = 5.23 × 10-7, R2 = 0.24, MAPE = 1.61) while providing reliable uncertainty quantification (PICP = 100%, MPIW = 0.004154). Despite high sensitivity (80%) and moderate specificity (75%), the model effectively captured key temporal-spatial patterns of Salmonella occurrence. Key predictors included overall food contamination prevalence, poultry products, raw livestock and poultry meat, and fish. Integration of food contamination data substantially improved predictive accuracy and interpretability, demonstrating the potential of QRF for uncertainty-aware risk assessment and early warning within food safety surveillance systems. Overall, this framework underscores the value of probabilistic machine learning and multisectoral data integration for proactive control of foodborne risk and prioritization of resources in food safety management.