| First, a feedforward error back-propagation neural network (FEBNN) model (classifier) was developed to predict survival and growth of Escherichia coli O157:H7 in response to five environmental conditions (temperature, pH, and concentrations of acetic acid, sucrose and salt). The neural network was trained by using a data set from controlled experiments conducted with a cocktail of five strains of E. coli O157:H7 in tryptic soy broth. It correctly predicted the growth/no-growth in 1810 (99.5%) with 8 false positives and 2 false negatives, and survival/death in 1804 (99.1%) with 13 false positives and 3 false negatives. Thirty data from experimental mayonnaise inoculated with E. coli O157:H7 and two literature data sets (26 conditions) were used for experimental validation. The FEBNN model predicted the survival/death in 27 of 30 cases (90.0% accuracy) with three fail-positive predictions and all observed growth (100%).; In addition, a knowledge-based neural network envelope (KBNNE) using surrogate pathogens was developed to characterize the uncertainty about the probability of infection at an ingested dose level. Eight E. coli O157:H7 outbreak data with fractional animal and surrogates data were combined to build the knowledge-based neural network dose-response (KBNNDR) model. A constant variation of approximate 95% confidence limits of the KBNNDR was made to interpret the uncertainty. A fuzzy rule-based model was developed to characterize the various uncertainty of E. coli O157:H7 dose-response, when the lack of data severely restricted the accuracies of beta-Poisson, KBNNE and KBNNDR models. The fuzzy model provided a set of fuzzy zones. Each fuzzy zone demonstrated the different width of response interval at different dose levels to describe the uncertainty, variability and imprecision of E. coli O157:H7 dose-response. |
