A model order selection criterion for the identification of physiologic systems

We propose a model order selection criterion for the identification of a linear regression model which can be an adequate representation of a resting physiologic system. The criterion, which is derived by estimating the mean squared parameter error weighted by the input data covariance matrix, is called WPE and reflects a trade-off between mean squared prediction error and model complexity. We compare the asymptotic performance of WPE with the widely used final prediction error (FPE). We also demonstrate through simulated and physiologic data that WPE minimization provides a more accurate and succinct characterization of system dynamics than FPE minimization. To our knowledge, WPE has not been previously proposed for model order selection.