Egg authentication under seasonal variation using stable isotope analysis combined with machine learning classification

In the European Union (EU) eggs from laying hens are sold in different husbandry systems: enriched cage, barn, free-range and organic. An analytical distinction between the husbandry systems is of interest for the food control authorities, as fraudsters can make high profits by mislabelling the husbandry system of the eggs. In this study, a total of 180 hen's egg samples labelled to originate from barn, free-range and organic hens were taken from food retailers in Germany at monthly intervals over a period of 15 months. The eggs were analysed using stable isotope ratio analysis (IRMS) in order to identify potential influences of different husbandry methods. The analysis on a monthly basis showed that the feeding of the laying hens varied over the year, which makes it difficult to make a clear classification according to the type of husbandry. Temporary events such as the officially ordered poultry quarantine due to the avian influenza, had minor influence on the differentiation of the husbandry types. Various statistical models including linear discriminant analysis (LDA), random forest (RF) and multinomial logistic regression (Multinomial) based on a multi-element approach (δ 13 C, δ 15 N, δ 34 S) were used to classify the eggs according to the three types of husbandry. On average, the models achieved a correct classification of approximately 70%. The reduction of the prediction model to two classes (organic vs. conventional eggs) led to a correct classification of almost 80% across all samples. Additionally, significant differences in the mean values of the sulphur isotopes of the organic eggs showed a regional differentiation between Bavaria (southern Germany) and Schleswig-Holstein (northern Germany). • An IRMS method to classify eggs by the husbandry system is described. • Monthly sampling reveals seasonal fluctuations in stable isotope values. • Classification accuracy is improved by modelling multi element data. • Periods of quarantine due to avian influenza have minor influence on classification.