Ensemble learning: a study on the border inspection prediction model for high-risk products adulterated with Sudan Red in Taiwan

Sudan Red (SR), an industrial azo dye classified as a Group 3 carcinogen, remains a food safety concern despite bans and Taiwan’s Border Prediction Intelligent System (BPIS). To address the lack of validated models for rare hazard detection, we developed an ensemble learning model combining Random Forest, Naïve Bayes, and a Multilayer Perceptron with synthetic oversampling. Trained on 3,699 inspection records and validated with 700 real cases, the model detected two of three SR-contaminated batches by sampling only 31 shipments (hit rate 6.45%, coverage 66.67%), outperforming BPIS (1/220; 0.45%, 33.33%). Statistical tests (risk difference +5.997%, relative risk 14.19, p = 0.041; Bayesian posterior 99.3%) confirmed superiority. Deep learning improved predictive robustness, and trade partnerships with importers were significantly linked to SR violations (χ 2 test, p <0.001). This hazard-specific, data-driven approach improves detection efficiency, reduces inspection burden, and provides an adaptive framework for risk-based border food safety control. • A predictive model was developed in Taiwan to identify Sudan Red residues in food. • Random Forest, Naive Bayes, and deep learning are integrated into the model. • The new model doubles the coverage of non-compliant batch predictions. • The model reduces the number of batches requiring sampling for inspection. • Correlation between importer partnerships and Sudan Red residue presence was found.