Abstracts

Rationale: Intracranial EEG (ICEEG) provides considerable information about ictal and interictal activity . Advanced signal processing techniques can provide tools for better determination of the seizure onset and possibly seizure prediction. Modern single channel methods allow for analysis of epileptic activity in each channel to characterize the non-stationary dynamics of seizure development and propagation. Methods based on autoregressive models describe linear interactions between different brain regions and can provide insight into neural network functional connectivity, but describes only the stationary parts of the signal. Usually these methods are applied in isolation. By combining them one can take advantage of the strength of each method and perform more comprehensive analyzes of seizure dynamics, particularly of high-frequency components. Methods: ICEEG recordings from eighteen consecutive patients with subdural grids and/or strips were analyzed for this study. Eleven patients were diagnosed with mesial temporal lobe epilepsy and seven with neocortical onset epilepsy. Seizures identified during monitoring were initially analyzed with the matching pursuit (MP) time-frequency decomposition to identify the 10-15 channels most active during time of seizure onset. Events included sub-clinical seizures and partial seizures (simple or complex) with or without secondary generalization. The signals from channels selected were fitted with a multichannel autoregressive model. The residuals after autoregressive modeling were then decomposed into Gabor atoms using the MP decomposition. Results: The multivariate autoregressive model best describes the stationary part of the signal. In ICEEG recordings usually only low frequency components can be treated as stationary during windows of analysis (1sec). Higher frequency transient components are not well described by autoregressive model and are left in the residue. The MP decomposition of the residue provide a better description of those transients. In analyzed data there were significantly more high frequency transients present in the MP decompositions of the residuals than in the original decompositions of the signal without autoregressive pre-processing. The time-frequency decomposition of the whole signal is dominated by high energy low frequency components which obscure the low amplitude high frequency transients. Conclusions: Pre-processing of the ICEEG with multivariate autoregressive models fitted to the signals increases sensitivity of the MP decomposition for high frequency transients. The number of relatively high energy high frequency components in the matching pursuit decomposition changes from higher during seizure onset to lower in postictal periods. Application of multivariate autoregressive models combined with MP decomposition allows for separation of linear interactions of stationary components of the signal from non-stationary transients providing more clear descriptions of seizure dynamics. Supported by NIH grant NS48222.