Introduction

The low amplitude, transient high frequency rhythms in intracranial EEG recordings can provide important information aiding detection and localization of seizure onset. Unfortunately most of the signal analysis methods tend to emphasize components of greater energy, i.e. components lasting longer and/or with larger amplitude. One solution is prefiltering data with arbitrary high-pass filter. This approach would also emphasize higher frequency harmonics of periodic rhythms of low frequency. The multichannel autoregressive (AR) model can be used to describe the periodic portion of the dynamics of the EEG signal during both ictal and interictal periods (Franaszczuk et al. 1994). The time-frequency analysis using the Matching Pursuit (MP) method provides a detailed description of the dynamics of the seizure, even during the most rapidly changing periods (Franaszczuk et al. 1998). In these analyses, however, the time-frequency decomposition is dominated by high energy, quasi-periodic, low frequency components. The low amplitude non-periodic (usually high-frequency components) tend to be underemphasized in MP decompositions. In this study we use an autoregressive multichannel model to remove most of the periodic components prior to MP analyses. This allows for a more thorough subsequent analysis of transient components using the MP method. The state space representation of a multichannel autoregressive model can be interpreted as a linear approximation of a nonlinear dynamic system (Franaszczuk and Bergey, 1999). In this case the residuals from AR method represent a non-linear and non-periodic portion of the system dynamics.