Abstracts

RATIONALE: One of the common features of physiological time-series is the abundance of noise that can mask or distort events such as interictal spikes and epileptic seizure onsets. This can hinder the automatic classification of epileptiform events. One way to reduce the influence of noise is to extract the essential information from a given type of events by recording a large series of those events and subsequently clustering them according to some distance measure. Such a procedure is meaningful only when the signals are defined in the same measurement frame where they can be compared and where those belonging to the same clusters can be averaged.
METHODS: We propose a new approach to the problem of classification of time events in multi-channel signal recordings. An essential phase of such a classification is the alignment of the different events, or in more general terms, the transformation of the data to a common reference time frame. The common reference frame was reconstructed applying time-translation based on delayed mutual correlation functions of the individual events. The proposed method is applicable to more complicated cases such as seizure onsets. To validate our technique and to compare it with the standard clustering techniques we used a signal to noise measure defined in each time point as the ratio between the channel-averaged standard deviation between the members of a given cluster and the inter-channel standard deviation of the cluster-averaged signal. In addition, we used a single moving dipole localisation method to compare the results. The method is applied to 151 channel magnetoencephalograph (MEG) data sets recorded from four epileptic patients showing epileptiform discharges: two patients had focal epilepsies, and another two had photosensitive epilepsy. One patient of the latter group had photically induced absences.
RESULTS: We were able to find 3 clusters of 19,14 and 12 out of total of 50 frontal spikes, 2 clusters of 8 and 9 out of 18 temporal spikes, one cluster of 6 out of 9 photo-induced occipital spike and waves discharges (SWD) and a cluster of 6 out of 11 onsets of photo-induced 3Hz spike and wave absence seizure. In all cases the merits of the proposed signal alignment paradigm were quantified by the signal to noise ratio of the corresponding clusters. In the case of the photically induced SWD, the traditional method failed to cluster simultaneously the spike and the slow wave components of the signal while the new technique succeeded. In the case of absence-seizure onsets, the alignment and clustering technique showed a common onset template. This last result was only possible to achieve with the new technique as the traditional, feature-based alignment could not be applied for complex events.
CONCLUSIONS: Our method represents an improvement relative to the usual clustering methods where signal alignment is based on the identification of some local feature. The quality of the classification is validated by the signal to noise ratio analysis. Dipole localisation solutions can be positively affected by our method for cases with plausible single-dipole source.
[Supported by: SEIN, Scientific Research, Heemstede, The Netherlands]