Development and Validation of an Explainable AI–Bayesian Framework for Global Food Safety Risk Prediction and Supply Chain Resilience

Abstract

This study developed and validated an explainable AI–Bayesian framework that predicts food safety incident risk in global import supply chains and links predicted risk to supply chain resilience. The framework fuses a tree-augmented Bayesian network with Dirichlet-smoothed conditional tables, a gradient-boosting model explained through SHAP attributions, and a sparse logistic model through Bayesian fusion weights with a Laplace posterior, producing calibrated incident probabilities with credible intervals and ranked drivers for each supply chain. Because harmonised multi-source panels with verified incidents are not available at global scale, the framework was validated on a simulation-calibrated network of 260 import supply chains across five origin regions and six product categories observed over 60 months, with incident propensities ordered to published burden, fraud and notification patterns; 240 supply chains were retained, yielding 10,800 supply-chain-months, 435 incidents and 374 disruption episodes. Audit score (β = −0.49), price pressure (β = 0.21), chain length (β = 0.18) and supplier concentration (β = 0.12) predicted incident rates (R² = 0.465, F(9, 230) = 11.79, p < .001). Under forward-chaining cross-validation the hybrid achieved a mean area under the curve of 0.808, Brier score of 0.078 and expected calibration error of 0.023, outperforming the historical-frequency rule (ΔAUC = 0.049, p = .042) and the Bayesian network alone (ΔAUC = 0.031, p = .020; ΔBrier = −0.022, p = .005), exceeding gradient boosting in precision–recall area (p = .038) and matching logistic regression; the Bayesian component retained discrimination when the laboratory stream was withheld (AUC 0.829 to 0.825) whereas boosting fell from 0.850 to 0.818. SHAP rankings were stable across folds (ρ = 0.73). A Resilience Capability Index (α = 0.85) predicted recovery time (β = −0.64) and service loss (β = −0.67) across disruption episodes (R² = 0.76 and 0.80), with recovery falling from 9.5 months in the lowest tertile to 4.9 in the highest. Bayesian fusion of explainable learners thus delivers calibrated, robust and interpretable risk prediction, and resilience capability, rather than predicted risk, governs recovery.