Use of survival analysis and Classification and Regression Trees to model the growth/no growth boundary of spoilage yeasts as affected by alcohol, pH, sucrose, sorbate and temperature

This paper describes the application of a survival analysis model and a Classification and Regression Trees (CART) model to a data set comprising times to growth of a yeast cocktail inoculated into media simulating a fruit-based or alcoholic food or drink, and covering over 900 combinations of five environmental factors (alcohol, pH, sucrose, sorbate and temperature). Growth was determined as either the time to growth within a 150-day time period or as no-growth after 150 days. Models were developed which could either predict the likelihood of growth occurring within the 150 day period, or the time to grow, either in days or in one of three categories chosen to represent a rapid (1–14 days), medium (15–30 days) or slow (31–150 days) growth response. was observed in 29% of the experimental conditions and demonstrated that the yeasts used were able to grow under extreme environmental conditions, for example at a pH value of 2.1, a temperature of 2 °C, a sucrose concentration of 55% (w/w) or an alcohol concentration of 12% (w/v). Generally, both models provided a reasonable fit to the data, and successfully predicted the growth class in 84% of cases. Direct comparisons of the models were made to determine the more suitable for predicting the growth of yeasts in food systems. The survival analysis model was preferred for this data set because it was more fail-safe than the CART model. d validation studies, the survival model generally gave reliable predictions of time to growth in a range of 23 different food and drink products and is considered to be a reliable model to predict the likelihood and speed of yeast spoilage for a range of fruit-based or alcoholic food or drinks.