Purpose: To develop models and software to examine the effects of different input parameters such as preservatives, food additives, carbonation and pH on the microbial stability of beverage formulations.
Methods: A database of experimental data involving 659 beverage formulations each characterized by seven explanatory variables was used to develop two separate models. Since the outcome variable is binary (product is either stable or unstable) a logistic regression model with interaction terms was fitted to the data. Various statistical methods and tests were used in order to select and validate the models including deviance, sensitivity and specificity, detection of outliers and influential observations.
Results: It was found that dividing the experimental data into two categories, carbonated and non-carbonated, and constructing two separate models for each case gave the best results in terms of optimizing the accuracy of predictions, giving an sensitivities of 85% and 82%, respectively. Given appropriate input parameters within the experimental range, the models predict whether a product is stable or not and outputs the probability associated with the prediction. The developed and validated models were integrated into a web-based software system which can be used for routine stability assessment.
Significance: The software tool presents a rational and realistic methodology for assessing the microbial stability of beverage formulations. This helps the food manufacturers to more efficiently exploit the effects of changing input parameters on the stability and safety of their products without the requirement of performing as many challenge tests.