Abstract
The pre-ictal epileptiform discharges can hardly be distinguished from the scalp Electroencephalogram (EEG). However, detecting them from the intracranial EEG is much easier. On the other hand, the drawback of using intracranial EEG is the invasive insertion of electrodes into the brain's tissue. Furthermore, for recording these signals patients have to be in hospital under anaesthesia. Conversely, scalp EEG can be recorded easily from normal people as well as patient awaiting surgery under much simpler circumstances. In this work the focus is on extraction of pre-ictal epileptiform discharges from scalp EEG by developing a suitable blind source separation (BSS) algorithm. The first proposed method is based on creating a template from intracranial data, which is then used in the form of a constraint in a BSS algorithm. To generate a suitable template, the segments during which the brain discharges are labelled are used to generate the necessary templates. Approximate entropy (ApEn) is used for detecting these segments using chaoticity measurement. Constrained BSS using independent component analysis (ICA) is then applied to the scalp data to extract the desired source and to evaluate its effect on scalp electrodes. The effectiveness of such a constrained approach has been demonstrated by comparing its outcome with that of the unconstrained method. Such a BSS model can be later directly used to separate the desired components from the scalp EEG only. Exploiting sparsity is known to be very beneficial in BSS. Even if data is not sparse in its current domain, it can be modelled as sparse linear combinations of atoms of a chosen dictionary. This brings up the idea of having a dictionary of the waveforms for almost all possible epileptic discharges. This is a generalisation of the previous method by increasing the number of template from one template to a number of them. The suitable combination of atoms which would best approximate the source of interest are determined. This forms the basis of the second part of the work. The results of applying this method have been compared with that of the conventional BSS. In the final part of the work, the dictionary is partly pre-specified based on chirplet modelling of various kinds of real epileptic discharges, and partly learned using a dictionary learning algorithm. The learned part is added to account for other sources which are are present in the EEG as well as the desired epileptic discharges. The dictionary which includes a fixed and a variable (i.e. learned) part, is incorporated into a source separation framework to extract the closest source to the source of interest from the mixtures. Experiments on synthetic mixtures of real data consisting of epileptic discharges, and on real scalp recordings are used to evaluate the proposed methods, and the results are compared with those of traditional BSS algorithms. In addition, the importance and significance of the method has been tested for a number of cases to validate it for clinical purposes. The detection rate of the proposed method is compared to the clinician's scorings from scalp EEG, which shows a significant improvement. Applying the proposed method helps reduce the number of invasive brain insertions for recording purposes prior to surgery.