AI-Driven Early Detection of Global Foodborne and Zoonotic Disease Threats: A Spatiotemporal Machine Learning Approach

Abstract

This study developed and evaluated a spatiotemporal machine learning framework for the early detection of foodborne and zoonotic disease threats, integrating official case reporting, event-based media intelligence, animal-health and food-chain monitoring, climate anomalies, and mobility-weighted neighbourhood signals into a single region-week early-warning pipeline. Because harmonised multi-stream surveillance panels with verified onset dates are not available at global scale, the framework was validated on a simulation-calibrated network of 192 regions in five macro-regions observed over 260 weeks, with outbreak hazards ordered to published regional burden estimates and zoonotic emergence correlates; 180 regions were retained, yielding 29,199 at-risk region-weeks and 734 outbreak episodes, 44.4% of them zoonotic. Regional vulnerability covariates explained 42.1% of the variance in outbreak incidence (R² = 0.421, F(6, 173) = 11.21, p < .001), with surveillance capacity (β = −0.28), connectivity (β = 0.27), livestock density (β = 0.23), and climate-anomaly exposure (β = 0.18) as significant predictors. Under forward-chaining cross-validation, a spatiotemporal multi-source gradient-boosting model achieved a mean area under the curve of 0.800 against 0.606 for a model restricted to official counts (Δ = 0.194, corrected t = 9.59, p < .001) and 0.474 for a conventional control chart; sensitivity at a 5% false-alarm rate rose from 8.4% to 45.9%, and the share of episodes flagged within four weeks before onset from 13.1% to 60.8%. Spatial neighbourhood and regional attributes added 0.034 to the area under the curve (p = .055), while gradient boosting and logistic regression performed equivalently on the same information set. A Signal Convergence Index built from three surveillance streams (α = 0.65) correlated with subsequent case surges (r = 0.21, p < .001) and with detection lead time (ρ = 0.37, p < .001); episodes in the top convergence tertile were flagged before onset in 85.0% of cases compared with 38.3% in the bottom tertile. The findings indicate that the information set, rather than the algorithm, drives early detection, and that a spatiotemporal One Health signal architecture can convert routinely available data into actionable early warning.