Separation of spiky transients in EEG/MEG using morphological filters in multi-resolution analysis

Epileptic electroencephalographic (EEG) data often contains a large number of sharp spiky transient patterns which are diagnostically important. Background activity is the EEG activity representing the normal pattern from the brain. Transient activity manifests itself as any non-structured sharp wave with dynamically short appearance as distinguished from the background EEG. Generally speaking, the amplitude change of background activity varies slowly with time and spiky transient activity varies quickly with pointed peaks.; In this thesis, a method has been developed to automatically extract transient patterns based on morphological filtering in multiresolution representation. Using a simple structuring element (SE) to match a signal's geometrical shape, mathematical morphology is applied to detect the differences of morphological characteristics of signals. If a signal contains features consistent with the geometrical feature of the structuring element, a morphological filter can recognize and extract the signal of interest. The multiresolution scheme can be based on the wavelet packet transform which decomposes a signal into scaling and wavelet coefficients of different resolutions. The morphological separation filter is applied to these coefficients to produce two subsets of coefficients for each coefficient sequence: one representing the background activity and the other representing the transients. These subsets of coefficients are processed by the inverse wavelet transform to obtain the transient component and the background component. Alternatively, a morphological lifting scheme has been proposed for separation these two components. Experimental results on both synthetic data and real EEG data have shown that the developed methods are highly effective in automatic extraction of spiky transients in the epileptic EEG data.; The interictal spike trains thus extracted from multiple electrode recordings are further analyzed. Their cross-correlograms are examined according to the stochastic point process model. Our experiment result has been verified by human experts' estimation.