Comparing three differing approaches to identify a three-parameter gas-exchange model with noisy data

Abstract

Using mathematical models enables simulation of patient individual physiology. It can therefore be employed for predicting the outcome of various therapy settings. To be able to utilize a model at the bedside it has to be identifiable using the available data in a reasonable time. A previously presented identification approach that exploits hierarchical dependencies between models and that is independent of initial parameter estimates showed promising results. The presented work investigates how this approach behaves when the presented patient data is noisy. The method was evaluated employing data of twelve in-silico patients where noise of different amplitude was added. The results were compared to two alternative parameter identification approaches. One being the conventional method of identifying the model directly and the other being a method that iteratively reduces the dimension of the objective surface to optimize convergence (DRM – Dimensional Reduction Method). Both require a set of initial estimates which were taken arbitrarily from an increasing region around the true parameter values. Results show that the direct approach leads to a lower prediction error than both the hierarchical approach and the DRM when the initial estimates are close to the parameter values used to create the data, they become higher than the prediction error produced by the model identified with the hierarchical approach and the DRM when the initial estimates are drawn from a wider range around the true model parameters. Additionally, compared to the direct approach the DRM shows to be affected less by the initial estimates as shown by a more constant prediction error with respect to the region from which the initial estimates were drawn.