Use of ComBase data to develop an artificial neural network model for nonthermal inactivation of Campylobacter jejuni in milk and beef and evaluation of model performance and data completeness using the acceptable prediction zones method

Abstract

AbstractComBase is a widely used microbial modeling database. ComBase data can be used to develop and validate models and to test novel modeling methods like artificial neural networks (ANN) and acceptable prediction zones (APZ), which have been shown to outperform traditional methods. Here, ComBase data were used to evaluate the ANN and APZ methods for modeling nonthermal inactivation of Campylobacter jejuni in milk and beef as a function of time, temperature (−20, 1, 10, 20, 30, and 40°C), and strain (18177, ATCC 29428). Four ANN were developed using Excel and NeuralTools, and the best‐performing was a general regression neural network (GRNN) whose performance and data completeness were evaluated using the APZ method. Relative variable impacts in the GRNN model were 42.5%, 31.5%, 20.1%, and 5.9% for time, temperature, food, and strain, respectively. Nonthermal inactivation of C. jejuni was faster and greater at ambient than at cold temperatures and in milk than in beef except at 1°C where it was similar. The proportion of residuals in the APZ (pAPZ) ranged from 0.77 to 1 for individual nonthermal inactivation curves. Although the model had acceptable performance (pAPZ ≥0.7), it failed validation because of data gaps like one instead of four replicates per combination of independent variables and no data at −10°C. Thus, these and other data gaps identified need to be filled before the model can be used with confidence to predict behavior of C. jejuni in milk and beef. Nonetheless, results indicated that ANN and APZ methods can be used to model data for nonthermal inactivation of C. jejuni in food.