100A-32 |
Comparison of logistic regression and multivariate adaptive regression splines in modeling the effects of water activity, pH and potassium sorbate on growth/no growth of Saccharomyces cerevisiae |
R. XIONG and J. F. Meullenet. Food Science Department, University of Arkansas, 2650 N Young Ave, Fayetteville, AR 72704 Multivariate Adaptive Regression Splines (MARS), developed by Jerome Friedman, is an alternative to logistic regression for modeling bacterial growth/no growth interface because it can automate both variable and model selections. It is competitive with neural networks. To our best knowledge, MARS is such a new modeling tool to food scientists and has not yet been applied into food science including predictive microbiology. The objectives were (1) to apply the novel MARS modeling technique into predictive microbiology and (2) to compare it with logistic regression in modeling of the effects of aw, pH and potassium sorbate on growth/no growth of Saccharomyces cerevisiae. The published data set for growth/no growth of S. cerevisiae was collected. Effects of aw (0.97, 0.97, 0.93), pH (6.0, 5.0, 4.0, 3.0) and potassium sorbate (KS) (0, 50, 100, 200, 500, 1000 ppm) on growth of one strain of S. cerevisiae in laboratory media at 27oC were examined at selected incubation time periods (50 and 350 h). Any increase of the initially inoculated yeast count was considered growth and denoted as 1, while growth decline and no appreciable change in the initial inocula were considered no growth and denoted as 0. The data was modeled using both logistic and MARS models. For logistic model, the rates of overall correct prediction for 50 and 350 h were 91.43% (p<0.05) and 88.14% (p<0.05), respectively. For MARS models, the rates for 50 and 350 h were 98.75% (p<0.05) and 95.71% (p<0.05), respectively. These results show that MARS models were better than logistic models. In addition, graphical examination of the growth/no growth interface is much easier using MARS models. This implies that the MARS has potential uses in predictive microbiology as well as other food science problems dealing with recognizing underlying patterns in data.
Session 100A, Food Microbiology: General II
|