Application of Bayesian network modelling to predict food fraud products from China

Abstract

Previous studies identified that the origin of food fraud imported into EU was most commonly China. However, which food and drink categories were most affected by fraud? This study aimed to use bayesian network to predict food fraud products originating from China. RASFF notifications related to China were reviewed from 2004 to 2018.1668 fraud-related notifications were included in the development of the BN model. GeNie was used to construct the Bayesian network structure diagram and fraud risk category was directly linked to 6 of the explanatory variables: food and drink categories, year, hazard/others, notification by, origin or distributed via and action taken. The types of food fraud were divided into artificial enhancement (AE), adulteration, documentation, illegal trade, other and unauthorised activities. The BN model predicted the distribution of probabilities for food fraud type as AE (43.77%), other forms of fraud (20.20%), adulteration (15.95%), documentation (10.49), illegal trade (6.47%) and unauthorised activities (3.18%). Cereals and bakery products (21.34%) were most commonly affected by other forms of fraud (e.g. use of unauthorised genetically modified organism Bt63 in rice products) and adulteration with melamine and aluminium. Fruits and vegetables (12.65%) were affected by artificial enhancement (i.e. use of unauthorised pesticides or pesticide levels were above the maximum residue level). The model predicted 85% of the fraud correctly. The model is beneficial to border controls and inspections to select targeted food products for sampling and can be used to predict types of food fraud and the relationships between the variables.