Improving Microbial Growth Prediction by Product Unit Neural Networks

Authors

  • Cesar Hervás-Martíanez,

    Corresponding author
    1. Author Hervás-Martínez is with Dept. of Computing, Univ. of Córdoba, Campus Univ. de Rabanales, 14014 Córdoba, Spain.
    Search for more papers by this author
  • Rosa M. Garcíaa-Gimeno,

    1. Authors García-Gimeno and Zurera-Cosano are with Dept. of Food Science and Technology. Univ. of Córdoba, Campus Univ. de Rabanales, Córdoba, Spain.
    Search for more papers by this author
  • Alfonso C. Martíanez-Estudillo,

    1. Authors Martínez-Estudillo and Martínez-Estudillo are with ETEA, Dept. of Applied Mathematics, Univ. of Cordoba, Córdoba, Spain.
    Search for more papers by this author
  • Francisco J. Martíanez-Estudillo,

    1. Authors Martínez-Estudillo and Martínez-Estudillo are with ETEA, Dept. of Applied Mathematics, Univ. of Cordoba, Córdoba, Spain.
    Search for more papers by this author
  • Gonzalo Zurera-Cosano

    1. Authors García-Gimeno and Zurera-Cosano are with Dept. of Food Science and Technology. Univ. of Córdoba, Campus Univ. de Rabanales, Córdoba, Spain.
    Search for more papers by this author

Direct inquiries to author Hervás-Martínez (E-mail: chervas@uco.es).

ABSTRACT

This article presents a new approach to the Artificial Neural Networks (ANN) modeling of bacterial growth; using Neural Network models based on Product Units (PUNN) instead of on sigmoidal units (multilayer perceptron type [MLP]) of kinetic parameters (lag-time, growth rate, and maximum population density) of Leuconostoc mesenteroides and those factors affecting their growth such as storage temperature, pH, NaCl, and NaNO2 concentrations under anaerobic conditions. To enable the best degree of interpretability, a series of simple rules to simplify the expression of the model were set up. The new model PUNN was compared with Response Surface (RS) and MLP estimations developed previously. Standard Estimation Error of generalization (SEPG’) values obtained by PUNN were lower for lag-time and growth rate but higher for maximum population density than MLP when validated against a new data set. In all cases, bias factors (Bf) and accuracy factors (Af) were close to unity, which indicates a good fit between the observations and predictions for the 3 models. In our study, PUNN and MLP models were more complex than the RS models, especially in the case of the growth rate parameter, but they described lower SEPG’. With this work we have attempted to propose a new approach to neural networks estimations for its application on predictive microbiology, searching for models with easier interpretation and with a great ability to fit the data on the boundaries of variables range. We consider that still there is a lot left to do but PUNN could be a very valuable instrument for mathematical modeling.

Ancillary