Effectiveness of machine learning algorithms as a tool to meat traceability system. A case study to classify Spanish Mediterranean lamb carcasses

Abstract

Establishing the traceability of meat products has been a major focus of food science in recent decades. In this context, recent advances in food nutritional biomarker identification and improvements in statistical technology have allowed for more accurate identification and classification of food products. Moreover, artificial intelligence has now provided a new opportunity for optimizing existing methods to identify animal products. This study presents a comparative analysis of the effectiveness of different machine learning algorithms based on raw data from analyses of organoleptic, sensory and nutritional meat traits to differentiate categories of commercial lamb from an indigenous Spanish breed (Mallorquina breed) obtained from the following production systems: suckling lambs; light lambs from grazing; and light lambs from grazing supplemented with grain. Six machine learning algorithms were evaluated: Artificial Neural Network (ANN), Decision Tree, K-Nearest Neighbours (KNN), Naive Bayes, Multinomial Logistic Regression, and Support Vector Machine (SVM). For each algorithm, we tested three datasets, namely organoleptic traits and sensorial traits (CIELAB colour, water holding capacity, Warner-Bratzler shear force, volatile compounds and trained tasters), and nutritional traits (proximate composition and fatty acid profile). We also tested a combination of all three datasets. All the data were combined into a dataset with 144 variables resulting from the meat characterization, which included 11,232 event records. The ANN algorithm stood out for its high score with each of the three datasets used. In fact, we obtained an overall accuracy of 0.88, 0.83, and 0.88 for the organoleptic-sensory, nutritional, and combined datasets, respectively. The effectiveness of using the SVM algorithm to assign categories of lambs according to its production system performed better with nutritional traits and the full characterization, with performances equal to those obtained with ANN. The KNN algorithm showed the worst performance, with overall accuracies of 0.54 or lower for each of the datasets used. The results of this study demonstrate that machine learning is a useful tool for classifying commercial lamb carcasses. In fact, the ANN and SVM algorithms could be proposed as tools for differentiating categories of lamb production based on the organoleptic, sensory and nutritional characteristics of Mediterranean light lambs' meat. However, in order to improve the traceability methods of lamb meat production systems as a guarantee for consumers and to improve the learning processes used by these algorithms, more studies along these lines with other lamb breeds are required.