Parameter estimation combined with model reduction techniques for identifiability analysis of biological models

Abstract

Parameter estimation is typically used as part of model development to determine the values of unknown parameters. However, depending on the model complexity the number of parameters can also vary. High complexity models have large numbers of parameters requiring more computational effort to determine them and are also prone to overfitting. Low complexity models have smaller numbers of parameters but may have reduced accuracy. Based on available experimental data cross-validation can be used to compare different complexity models and determine the most appropriate complexity (James et al., 2013). Alternatively, it is possible to look at the identifiability of parameters based on experimental data which considers the sensitivity and correlation between parameters. Both these types of methodologies can be used to reduce the complexity of models such that insensitive and/or dependent/correlated can be removed or re-estimated and an alternative set of parameters can be computed. In this work both types of methods are explored with examples. Cross-validation combined with a Least Absolute Shrinkage and Selection Operator (LASSO) regularisation method is used to reduce the complexity of linear empirical equations for predicting the performance of downdraft biomass gasification (Binns and Ayub, 2021). Sensitivity and identifiability methods utilizing the Fischer Information Matrix (FIM) are used to reduce the complexity of a nonlinear system of partial integral differential equations describing a population balance model for microalgae cultivation (Usai et al., 2022). Application of these methods allows the number of parameters to be reduced depending on the tolerance and/or accuracy required. © 2023 Elsevier B.V.