Reliable food safety risk assessment via calibrated Mondrian conformal prediction: a machine learning framework with uncertainty quantification

Abstract

Food safety risk assessment requires predictive models that not only achieve high accuracy but also provide reliable uncertainty quantification to support regulatory decision-making. This study presents a comprehensive machine learning framework that integrates Mondrian conformal prediction with post hoc probability calibration for food safety risk classification using the Rapid Alert System for Food and Feed (RASFF) database. The task is framed as a binary classification problem, predicting whether an RASFF notification is serious or not. The proposed framework addresses three critical challenges in food safety prediction: class imbalance handling through the synthetic minority oversampling technique (SMOTE) combined with class weighting, probability miscalibration correction via Platt scaling, and distribution-free uncertainty quantification through group-conditional conformal prediction. The framework was evaluated on 22,643 RASFF notifications from 2019 to 2025 using a chronological training/calibration/test split. The eXtreme Gradient Boosting (XGBoost)-based framework achieves an accuracy rate of 0.784, a Cohen’s kappa of 0.296, and a Matthews correlation coefficient of 0.297. The Mondrian conformal predictor attains 90.3% overall coverage, with an average prediction set size of 1.216, while Platt scaling reduces the expected calibration error by 84%, from 0.267 to 0.043. A multitrial resampling analysis (5,000 random calibration/test splits) confirms that mean per-region coverage converges to the nominal 90% level, while single-split deviations in the smallest regions reflect finite-sample variance. Counterfactual explanation generation additionally provides actionable insights into risk factor modifications. The framework offers a decision-support prototype for food safety risk assessment that couples calibrated probabilities with distribution-free uncertainty quantification. Given the moderate agreement beyond chance and the reduced minority-class performance, the framework is positioned as a decision-support prototype requiring further validation before regulatory deployment. If restructuring is not preferred by the production team, the existing abstract may be retained as-is; the content of both versions is equivalent.