Modeling multivariate covariance nonstationary time series and their dependency structure: An application to human epileptic event EEG analysis

Abstract

The parametric modeing of covariance nonstationary time series and the computation of their changing interdependency structure from the fitted model are treated. The nonstationary time series are modeled by a multivaraiate time varying autoregressive (AR) model. The time evolution of the AR parameters is expressed as linear combinations of discrete Legendre orthogonal polynomial functions of time. The model is fitted by a Householder transformation-Akaike AIC method. The computation of the instantaneous dependence, feedback and causality structure of the time series from the fitted model, is discussed. An example of the modeling and determination of instantaneous causlity in a human implanted electrode seizure event EEG is shown.