Explainable AI for food safety risk prediction with uncertainty quantification using RASFF data

Food safety incidents pose significant public health risks globally, necessitating robust predictive frameworks for identifying high-risk contamination notifications. This study presents an explainable artificial intelligence framework for food safety risk prediction that integrates SHAP-based feature attribution, temporal precedence analysis, what-if sensitivity modeling, and conformal prediction for uncertainty quantification. The framework was developed and evaluated using 25,615 notifications from the European Rapid Alert System for Food and Feed (RASFF) spanning 2019–2025, with a strict 80/20 temporal train–test split. Among the evaluated classifier configurations, LightGBM achieved the highest AUC-ROC (the primary metric) on the final 18-feature set identified through ablation (AUC-ROC = 0.8970 [0.8884, 0.9052], Accuracy = 0.8087 [0.7976, 0.8196], F1 = 0.8077). Notification Type was the most influential predictor (mean absolute SHAP = 1.86), but it is a downstream regulatory variable available only after regulatory classification. We therefore distinguish a post-notification prioritization model that retains Notification Type (AUC-ROC = 0.8970) from a deployment-oriented early-prediction model that excludes it (AUC-ROC = 0.7993). The latter value, approximately 0.80, is the relevant estimate for early-warning deployment; the higher value applies only to post-notification prioritization. Historical-rate and count-statistic features were non-contributory after leakage correction (Delta-AUC = + 0.021 when removed). Granger temporal precedence testing with Augmented Dickey–Fuller (ADF) stationarity checks and Bonferroni correction ( p < 0.00167) identified three significant inter-hazard temporal relationships out of 30 tests (strongest: Pathogenic Microorganisms to Migration, F = 17.91). What-if sensitivity analysis produced a maximum model-predicted shift of 3.90 percentage points absolute (5.89% relative) under the EU-harmonization scenario; this is a feature-perturbation result rather than a causal policy-effect estimate. Multi-target prediction achieved hazard-type accuracy of 59.63% and notification-type accuracy of 65.02%. Conformal prediction with temporal calibration achieved 87.53% empirical coverage at a nominal 90% level; because the observations are temporally ordered, this result is reported as empirical rather than guaranteed coverage. All principal classification results are accompanied by bootstrap 95% confidence intervals and paired significance testing.