Explainable machine learning framework for foodborne disease outbreak prediction in Eastern Province, Saudi Arabia: case study

Abstract

The occurrence of foodborne diseases is a considerable public health issue, especially in areas that are quickly becoming urbanized with intricate food delivery systems. In this paper, we present a machine learning-based model for predicting outbreaks, explainability, and spatial risk propagation, validated through a multiyear data set of an epidemiological nature from 12 cities in the Eastern Province of Saudi Arabia (2021–2025). The final data set includes 61 cases and 13 engineered features. In the current research, the proposed architecture uses XGBoost to predict outbreaks, alongside using the random forest for predicting severity and support vector machine (SVM) for comparisons. The XGBoost classifier demonstrates an evenly balanced performance (accuracy = 0.85, precision = 0.78, recall = 0.78) on the testing set. Due to the size of the dataset, the results are provided with the estimation of uncertainty (rather than the exact numbers). Using leakage-safe repeated stratified cross-validation, the mean AUC equals 0.64 [95% interval = (0.20, 1.00)], and the leave-one-year-out validation method is not stable (mean AUC 0.47). Differences between the models (e.g., better single-split cross-validation AUC for SVM) are within confidence intervals. Interpretability is improved by using the SHAP framework to measure feature importance, which shows that the main factors are hospitalization and the severity of symptoms. The graph module also helps in understanding the propagation of disease risk between cities, highlighting the importance of well-connected metropolitan areas as disease hubs. Moreover, the use of a locally deployed Mistral LLM makes the generated explanations more readable. The findings show that our approach presents an appropriate balance of predictiveness, interpretability, and spatial knowledge. With only 61 data points and 11 outbreaks reported, this research is clearly not meant to be an early warning system, but rather a proof-of-concept on how one might be designed. In order to ensure reproducibility, the preprocessing pipeline and synthetic dataset generator have been made available to the community.