Quality Monitoring and Prediction of Vacuum-Packed Beef under Dynamic Cold Chain Conditions Using an Electronic Nose and Explainable AI-Based Regression Models

This is the authors' abstract. We don't add key points for this paper.

Temperature fluctuations during cold-chain distribution accelerate microbial growth and biochemical deterioration in beef, while temperature monitoring does not directly indicate the product’s actual quality. Besides, many machine learning models provide limited interpretation of the VOCs underlying their predictions. This study developed an interpretable electronic nose framework to predict vacuum-packed beef quality under constant and fluctuating temperatures. Beef samples were stored at 0, 5, 15, and 15 °C and under two fluctuating temperature profiles (FT1 and FT2). Physicochemical properties, total viable count (TVC), coliform count (CC), and VOC profiles of beef were monitored until the final experiment. Six regression algorithms were compared for predicting volatile basic nitrogen (VBN), TVC, and CC. The models were developed from constant-temperature data using a 70:30 (training: testing) split and k-fold grouped cross-validation, and used FT1 and FT2 for external validation. SHapley Additive exPlanations (SHAP) analysis was used to interpret VOC contributions in the optimal models. The results show that elevated and fluctuating temperatures accelerated physicochemical deterioration, microbial growth, and VOC changes; chromatographic analysis, PCA, and k-means clustering differentiated samples by storage progression and temperature exposure. AdaBoost performed the best prediction for VBN (R 2 = 0.910), whereas Gradient Boosting performed best for TVC (R 2 = 0.822) and CC (R 2 = 0.863). SHAP identified 3-(methylthio)propanal as an important predictor of VBN and 2-methylpropanal as the principal predictor of TVC and CC. Integrating electronic nose analysis, regression machine learning, external fluctuating temperatures validation, and XAI provides a promising framework for interpretable beef quality monitoring and intelligent cold-chain distribution.