Explainable machine learning framework for foodborne disease outbreak prediction in Eastern Province, Saudi Arabia: case study
Frontiers in Public Health·
- DOI
- 10.3389/fpubh.2026.1883871
- PMID
- —
- PMCID
- —
- OpenAlex
- —
- Study type
- Journal article
- Publisher
- Frontiers Media SA
- Article type
- journal-article
- Integrity
- current
Why this research matters now
The authors frame the work as a proof-of-concept illustrating how predictive modeling, interpretability, and spatial risk mapping could be combined to support future foodborne outbreak surveillance in urbanizing settings.
Structured evidence summary
Research question
The study explores whether a machine learning framework can predict foodborne disease outbreaks, characterize severity, and illustrate spatial risk propagation across cities in Eastern Province, Saudi Arabia.
Study design
A retrospective multi-year epidemiological case study using a 61-case, 13-feature dataset from 12 cities (2021–2025). The modeling pipeline combines XGBoost for outbreak prediction, random forest for severity prediction, and SVM as a comparator, with SHAP for interpretability and a graph module for spatial risk propagation; validation uses leakage-safe repeated stratified cross-validation and leave-one-year-out splits, with results reported with uncertainty estimates.
Population and setting
The setting is 12 cities in the Eastern Province of Saudi Arabia, drawing on epidemiological records from 2021–2025 with 61 cases and 11 outbreaks.
Main findings
XGBoost achieved an accuracy of 0.85, precision of 0.78, and recall of 0.78 on the test split, with a mean AUC of 0.64 (95% interval 0.20–1.00) under leakage-safe repeated stratified cross-validation; leave-one-year-out validation was unstable (mean AUC 0.47). Differences between models, including a better single-split cross-validation AUC for SVM, fell within confidence intervals. SHAP identified hospitalization and symptom severity as the main contributors, and a graph module highlighted well-connected metropolitan areas as disease hubs.
Public-health relevance
The authors frame the work as a proof-of-concept illustrating how predictive modeling, interpretability, and spatial risk mapping could be combined to support future foodborne outbreak surveillance in urbanizing settings.
Important limitations
The authors explicitly note that with only 61 data points and 11 reported outbreaks the study is not intended as an early warning system but as a proof-of-concept; broad confidence intervals and unstable leave-one-year-out performance further indicate limited predictive robustness, and results are reported with uncertainty estimates rather than exact values. The supplied evidence here is limited to the article's abstract and metadata, and the original paper is required for decision-grade interpretation.
GIDS interpretation
Within GIDS, this record is discoverable as a Saudi Arabia-focused proof-of-concept on foodborne outbreak prediction methodology, useful for context on analytic approaches rather than as confirmation of an active surveillance signal.
Related GIDS surveillance
Literature context does not validate, explain, or change a surveillance signal. Exact and contextual relationships are shown separately.
Evidence relationships
This article has 5 auditable classifier relationships to diseases, places, topics, and study design.