Advancing MEG- and EEG-based decoding of motor imagery for practical brain-computer interfaces for neuro-rehabilitation

Abstract

Non-invasive electroneurophysiology technologies such as Electro- and Magneto-encephalography (EEG and MEG) provide efficient ways to capture neural patterns related to perceived actions. The thesis contributes towards making a more efficient and reliable Brain-computer interface (BCI) systems by providing enhanced methodologies for processing EEG/MEG data. Extensive BCI literature reviews have identified open research challenges in terms of the low performance of M/EEG based single-trial as well as continuous decoding systems due to the limited number of trials, intra-subject inter-session variability, and inter-subject information transfer variability. The research undertaken towards addressing these challenges has led to four original
research contributions. In MEG, 306 acquisition channels are used compared to low number of single session trials, i.e. 60-80. In the first contribution, a channel selection-based methodology is introduced to significantly improve classification performance by keeping only positively contributing channels towards classification. This study shows, selecting positively contributing channels resulted in significantly higher classification accuracy using band-power and common spatial patterns (CSP) as features, with much reduced computational load and processing time. The method reduces the number of channels significantly from 204 to a range of 125, using band-power as a feature and to 15-105 for CSP. Classification performance drops across sessions perhaps due to participants’ head movements and brain dynamics changes across sessions resulting in poor feature decoding by classifiers. Thus, the second contribution introduces a novel concept of Megablocks using CNN based architectures which can be used for inter-subject continuous decoding with better accuracy for developing calibration-free MI-BCIs. These Megablocks repeat a specific architecture block several times such as one or more convolutional layers in a single Megablock for adapting the network against inter-subject variabilities. However, classification using Megablocks uses CNN which requires a large amount of data for training the models. As EEG/MEG requires participants for data collection, it is difficult to collect large amounts of data as it is time consuming. Thus the third contribution introduces a method to generate artificial EEG motor imagery (MI) signals for the first time to counteract the low trial numbers in MI-based EEG-BCI. The research contributes to producing synthetic EEG samples. The enhanced BCI systems developed facilitated continuous decoding of trials with better performance for real-time interaction with the environment which resulted in the fourth contribution to develop the first MEG compatible exoskeleton for active assistance in undertaking therapeutic motor exercises helpful in functional brain recovery of upper limb impaired stroke patients.

Date of Award	Dec 2021
Original language	English
Sponsors	UKIERI & Department for the Economy
Supervisor	Girijesh Prasad (Supervisor) & Karl Mc Creadie (Supervisor)