A Multidimensional Stacking Ensemble Framework for Predicting Foodborne Disease Risks Originating from Animal Sources

<title>Abstract</title> <p>Foodborne diseases impose a major global public health burden, yet existing surveillance systems remain largely reactive. This study develops a multidimensional machine learning framework for predicting disease risks associated with animal products in Zhejiang Province, China. Drawing on 175,803 patient surveillance records collected between 2014 and 2023, supplemented by diverse meteorological, environmental, and socioeconomic indicators, we constructed a systematic feature engineering pipeline. We comprehensively evaluated seven predictive models, including algorithms based on decision trees and classical time series approaches. Ultimately, a hierarchical Stacking ensemble approach, integrating XGBoost, LightGBM, and Random Forest with a Ridge regression secondary algorithm, demonstrated superior predictive performance on the independent testing set. This optimal ensemble recorded a Mean Absolute Percentage Error of 11.30\%, alongside a Mean Absolute Error of 218.90 and a Root Mean Square Error of 273.65, significantly outperforming all individual baseline models. Furthermore, an integrated SHapley Additive exPlanations (SHAP) analysis identified three critical predictive dimensions: annual seasonality, spatial heterogeneity across prefectures, and specific food source contexts. Within these contexts, catering venue types and bulk processing methods emerged as the strongest predictive indicators. These findings definitively demonstrate that integrating advanced ensemble learning with explainable artificial intelligence can transition epidemiological surveillance from reactive monitoring to proactive risk stratification, providing a highly scalable template for proactive food safety governance.</p>