Purpose: To develop multiple versions of a neural network model for growth of Salmonella serotypes in ground chicken thigh meat for the purpose of providing the food industry and risk assessors with models that meet their needs for improving the safety of chicken.
Methods: An automated miniature most probable number (MPN) method was used to enumerate Salmonella serotypes (n = 8) in ground chicken thigh meat portions (0.75 cm3) during cold storage at 16°C for 0 to 8 days. A multiple-layer feedforward neural network model was developed using Excel and NeuralTools. Two additional versions of the Excel model were developed: a version for the food industry that does not require NeuralTools to run and a stochastic version for risk assessors that requires NeuralTools and @Risk to run. Performance of the model was evaluated using the acceptable prediction zone method.
Results: Growth of Salmonella was affected (P < 0.05) by serotype at 4, 6, and 8 days of storage but not at 0, 1, or 2 days of storage at 16°C. Maximum log MPN per portion ranged from 6.12 ± 0.47 (mean ± SD) for serotype 8,20:-:z6 to 6.84 ± 0.23 for serotype Thompson. The proportion of residuals in an acceptable prediction zone (pAPZ) from -1 log (fail-safe) to 0.5 log (fail-dangerous) was 0.948 for training data (n = 192) and 0.988 for testing data (n = 84). A pAPZ ≥ 0.7 indicates that the model provided predictions with acceptable bias and accuracy. Thus, the model was successfully validated.
Significance: The new models will allow the food industry and risk assessors to better assess the risk of salmonellosis from chicken subjected to temperature abuse during refrigerated storage.