Bayesian network as an aid for Food Chain Information use for meat inspection

Current ante mortem inspection involves a check of relevant Food Chain Information (FCI) transmitted by the farmer to the slaughterhouse on a regulatory FCI document. Since 2000, a farm sanitary form with FCI data has been used for all consignments of broiler chickens in France. However, the FCI needs to be standardized for the collection and interpretation of data. A study was conducted to develop an expert system, undertaken to elaborate on a simple decision support system capable of predicting whether the flocks will present a high condemnation risk, based on FCI. For this, 'optimal' (i.e. on-farm survey data) and 'worthy' (i.e. farmers' declaration on existing farm sanitary form) data quality conditions were considered to estimate the lower and upper reference bounds of the confidence that the decision-makers could have in such a tool. Chicken broiler flocks (404) were randomly selected in 15 slaughterhouses located in Western France in 2005. Condemnation proportion and farm sanitary form were collected for each selected flock. Information about health history and technical performances were also specifically collected on farm. Condemnation risk category was modelled from the on-farm collected information, using a Bayesian network and assuming this represented the optimal data quality conditions. Corresponding information declared by the farmer on the existing farm sanitary form was secondly used in the network to evaluate the impact of the uncertainty of such information on the condemnation classification obtained with the expert system. The learnt Bayesian network had 16 explanatory variables pertaining to technical characteristics and sanitary features of the flock. Using a threshold of 1% of condemned carcases to define high risk, the network sensitivity and specificity were 55% and 93%, respectively, corresponding to positive and negative predictive values of 70% and 87%. When declared existing information was used in the network, the sensitivity and specificity were 16% and 96%, respectively, corresponding to positive and negative predictive values of 57% and 80%. Results suggested that the predictive network developed may be insufficient for correctly classifying chicken flocks for targeting of management procedures, and in its current form, the expert system may be unlikely to be implemented in the field. However, it could help to improve the standardization of both form design and FCI interpretation at a national level.