Graph Convolutional Neural Network Approaches for Exploring and Discovering Brain Dynamics

Abstract

This thesis delves into Motor Imagery Electroencephalography (MI-EEG) classification, aiming to refine precision and deepen the understanding of the complex dynamics in motor imagery tasks. The study introduces five methodologies to enhance MI-EEG classification and investigate brain dynamics. The first, the Adjacency Convolutional Neural Network Model (Adj-CNNM), creates EEG graphs based on structural connectivity, improving subject independence and MI classification accuracy. The second, the Phase Locking Value Convolutional Neural Network Model (PLV-CNNM), focuses on functional connectivity, revealing intricate patterns during MI tasks and emphasizing connections within the frontal, central, and parietal brain regions. The ensemble methods—Stacked Ensemble Graph Convolutional Neural Network (S-Ensm-GCNN), Blend Ensemble Graph Convolutional Neural Network (Blend-Ensm-GCNN), and Vote Ensemble Graph Convolutional Neural Network (Vote-Ensm-GCNN)—integrate structural and functional connectivity, further optimizing MI classification accuracy across diverse patterns. Results highlight the effectiveness of these approaches, showcasing their superiority over state-of-the-art and baseline techniques. Specifically, the proposed methods reveal distinct activation patterns within the primary motor cortex (M1) and premotor cortex (PMC) specific to the imagined side of the body during motor imagery tasks. Foot movement tasks activate contralateral M1 and PMC, while tongue movement tasks engage the Supplementary Motor Area (SMA). These findings provide a more nuanced understanding of the neural dynamics underlying motor imagery, demonstrating the potential for improved classification accuracy in MI-EEG applications. This thesis contributes by introducing unconventional EEG graph representations, integrating multiple Graph Convolutional Neural Network (GCNN) models, and enabling diverse analyses for a comprehensive understanding of the structural and functional aspects of brain connectivity during MI tasks. Socially, the contributions extend to medical advancements, enhancing assistive technologies, promoting accessibility, improving healthcare efficiency, and inspiring future research.

This thesis not only advances MI-EEG classification methodologies but also provides profound insights into the dynamics of motor imagery tasks, with implications for neuroscience, neurorehabilitation, assistive technologies, and healthcare practices.