Seizure characterisation using frequency-dependent multivariate dynamics

The characterisation of epileptic seizures assists in the design of targeted pharmaceutical seizure prevention techniques and pre-surgical evaluations. In this paper, we expand on the recent use of multivariate techniques to study the cross-correlation dynamics between electroencephalographic (EEG) channels. The maximum overlap discrete wavelet transform (MODWT) is applied in order to separate the EEG channels into their underlying frequencies. The dynamics of the cross-correlation matrix between channels, at each frequency, are then analysed in terms of the eigenspectrum. By examination of the eigenspectrum, we show that it is possible to identify frequency-dependent changes in the correlation structure between channels which may be indicative of seizure activity. The technique is applied to EEG epileptiform data and the results indicate that the correlation dynamics vary over time and frequency, with larger correlations between channels at high frequencies. Additionally, a redistribution of wavelet energy is found, with increased fractional energy demonstrating the relative importance of high frequencies during seizures. Dynamical changes also occur in both correlation and energy at lower frequencies during seizures, suggesting that monitoring frequency-dependent correlation structure can characterise changes in EEG signals during these. Future work will involve the study of other large eigenvalues and inter-frequency correlations to determine additional seizure characteristics.