Event-related EEG analysis : $$b simple solutions of complex computations

Abstract

The value of EEG as a non-invasive technique for studying the time course and frequency composition of neuronal signals is well established. However, to date there is still no gold standard methodology for its analysis. Since the introduction of the technique many methodologies for artefact removal and signal isolation have been developed but their performance is often only assessed, against other methodologies, using simulated data with known and controlled artefacts and limited variance. Furthermore, these studies often only address a single stage in the entire analysis pipeline and do not consider the affect different preprocessing techniques might have upon the effectiveness of different signal analysis methodologies. To address this issue this thesis approaches the assessment of 4 different signal analysis methodologies using real-world-data, from two different stimulus evoked potential studies, and an EEG analysis pipeline that systematically applies and adjusts various preprocessing techniques before subsequent signal analysis. This semi-automated process can be broken down into two stages. Firstly, multiple configurations of a Preprocessing Optimisation Pipeline (POP) were performed to address three main causes of artefactual noise (1) electrical line noise, (2) non-neuronal potentials (low frequency drifts and muscle artefacts), and (3) ocular artefacts (blinks and saccades). Within the final stages of the POP data quality was assessed for each participant and poorly preprocessed participant datasets were excluded from further analysis based upon either a novel maximum baseline variability threshold criterion or a standard minimum epoch number threshold approach. Lastly, the data was passed onto a Signal Analysis Pipeline (SAP) which estimated the amplitude of task-specific signals of interest through one of four methodologies (1) grand average informed peak detection (GA-PD), (2) individual average peak detection (IAPD), (3) independent component analysis informed peak detection (ICA-PD) or (4) component of interest peak detection (COIPD). The effectiveness of each of the different preprocessing and signal analysis strategies were then assessed based upon observing the changes within task-specific outcome statistics.