Connectivity estimation of scalp electrodes after preprocessing. Application to seizure detection.

Different works using intracranial EEG recordings have demonstrated that the process of seizure generation is associated with a certain synchronization of groups of neurons. This synchronization can be estimated by parametric methods (AR model based coherence, directed transfer function, partial directed coherence) and non parametric methods (linear and nonlinear regression). However, despite the signals attenuation and noise or artefacts that contaminate traditional scalp EEG recordings, some works have found that is possible to detect this neural synchronization in scalp recordings. From here, various researches using scalp recordings have applied band pass filters in electrophysiological range or they even have excluded the noisiest electrodes as preprocessing method. The blind source separation (BSS) is a recently successful method for artefacts elimination. The goal of BSS is to recover the original sources (brain and artefactual sources), given only sensor observations. The artefactual sources are removed and the EEG is reconstructed with the brain sources only. In this communication we propose and evaluate a preprocessing method of real seizure EEG designed to improve connectivity estimation. The first step consists in source separation, followed by automatic source classification and “clean” EEG reconstruction. The reconstructed EEG is then used to estimate connectivity patterns. We show that the proposed technique improves the seizure detection, compared to similar connectivity methods applied without preprocessing.