Advancing computational analysis and modelling of EEG/MEG data

Abstract

To better understand the human brain, further advancements in the capabilities of computational modelling and analysis of functional neuroimaging data are required. Electro− and magneto− encephalography (EEG and MEG) technologies provide efficient ways to study the complex neural characteristics of both healthy and diseased brains. This thesis contributes towards extending the current approaches for analysis and functional connectivity (FC) modelling of M/EEG data. It includes an extensive review of the different computational modelling and analysis methods applied for the pre-processing, feature mining, and feature learning of the M/EEG signals in pursuit of single-trial neurophysiological patterns and their applications in various fields. This is followed by a review of current frameworks for modelling large scale functional brain networks. These reviews have identified open research challenges in terms of low performance of M/EEG based single-trial detection systems due to insufficient/inappropriate pre-processing and feature extraction methods and lack of neurophysiological validation of such systems. The research undertaken towards addressing these challenges have led to three original research contributions. In the first contribution, the estimation of current source density (CSD) is introduced as an essential pre-processing step for EEG analysis. It is shown that CSD significantly improves the distinction of motor-imagery (MI) related brain responses and has performed better than other referencing schemes (i.e. common reference and common average reference) and spherical surface Laplacian (SSL) methods (i.e. finite difference method and SSL using realistic head model). In the second contribution, EEG-based single-trial FC networks are introduced for MI (hand, feet, and tongue imagery kinaesthetic movements) and cognitive imagery (CI) tasks (word generation, mathematical subtraction, and spatial navigation) by implementing ‘partial’ granger causality modelling (PGCM) on two publically available EEG datasets. The outcome demonstrated that EEG brain networks for mixed imagery tasks (i.e. combination of CI and MI) can provide higher classification accuracy (for both binary and multiclass approaches) as compared to the current state-of-the-art method i.e. common spatial patterns (CSP). Moreover, it is shown for the first time that the FC between spatially distributed brain regions can provide additional useful discriminant information for the classification of the brain responses evoked during imagery tasks. In the third contribution, we investigated the temporal evolution and reorganisation of resting-state MEG FC networks over the four weeks of a multi-modal EEG-driven post-stroke upper limb (UL) movement rehabilitative therapy. The findings provide reliable brain connectivity patterns for evaluating UL functional recovery during stroke neurorehabilitation.

Date of Award	Feb 2019
Original language	English
Supervisor	Girijesh Prasad (Supervisor)