A Bayesian network to optimise sample size for food allergen monitoring

Generally, sampling size is optimised considering a single specific constraint. However, for financial reasons, only one sample is usually defined and used to satisfy several objectives. It is therefore crucial to choose a sample that meets all the required objectives. This paper proposes an original method for optimising a sample plan to monitor allergen traces in products consumed by allergy sufferers. The proposed method, based on a Bayesian network, enables several different constraints to be considered within a single model and the integration of literature data on concentration levels of allergen traces in food. Moreover, the construction of a three-stage sampling plan took into account the consumption preferences of peanut allergy sufferers between products with or without labels on the presence of allergen traces, and between the categories and subcategories of products. This method was applied to data from the MIRABEL project which aims to assess risks related to peanut traces for French allergy sufferers. The results show how the model used all the available information and constraints to balance the total number of samples set at 900 for food categories/subcategories and labelling types. As required, the model favoured the most consumed product categories and subcategories. At the same time, it increased the number of samples when peanut concentration is low. This helps reduce the uncertainty on peanut concentrations in these products and consequently on risk estimation. In conclusion, the proposed method is a useful tool for public administrations, risk assessors and risk managers to improve sampling plans for monitoring allergen traces or other health hazards in food. (C) 2014 Elsevier Ltd. All rights reserved.