Machine learning and food inspection: use of Bayesian Network modeling to support official controls in the food industries

This study aims to develop a machine learning model capable of predicting the type of non-compliance (NC) most likely to be detected by competent authorities during official control of food establishments based on their structural, product, and management characteristics. A Bayesian Network (BN) model was developed using data from 145 NCs detected by the Local Health Authority of Messina during 588 official controls performed on 101 approved food establishments between 2018 and 2021. The NCs were classified into 10 distinct categories based on the requirement not met: i) structural and equipment conditions; ii) water supply; iii) fight against pests; iv) hygiene of staff and processing; v) cleaning and sanitizing conditions; vi) raw materials, semi-finished and finished products; vii) labeling; viii) traceability; ix) hazard analysis and critical control points (HACCP); and x) microbiological criteria according to Regulation (EC) 2005/2073. The model was constructed by associating the number and type of NC with the criteria and corresponding evaluations established by the Veterinary Services for each food establishment risk categorization according to Annex 2 of the Intesa Stato-Regioni CSR 212/2016. In detail, 8 different criteria were considered: i) date of construction or renovation; ii) general maintenance conditions; iii) marketing area; iv) food category; v) product intended use; vi) professionalism of management; vii) hygienic-sanitary training of employees; and viii) HACCP self-control plan. The BN model was implemented using the Hugin Lite software, considering the NC type as the parent node and the 8 different criteria as the child nodes. The implemented model allowed the prediction of the most probable type of NCs by inputting the evaluations of each risk categorization criterion for a given food establishment into the child nodes. A total of 25 NCs detected in 10 food establishments during 2024 were used to validate the model. The validation cases were not included in the learning dataset. The model correctly predicted the occurrence of 19 NCs (76%), while 6 NCs (24%) were not predicted, and 3 NCs (12%) were rightly predicted as the most probable ones. Although further efforts are needed to implement the model with a greater amount of data, this study highlights the potential of a BN-based approach as a valuable tool for competent authorities in organizing and performing official controls from a new technological and sustainable perspective.