Accurately classifying electroencephalogram (EEG) signals, especially for individuals with neurodegenerative conditions such as myotrophic lateral sclerosis (ALS), poses a significant challenge due to high inter-subject and inter-session changes in signal. This study introduces a novel three-layer graph attention network (GAT) model for motor imagery (MI) classification, utilizing phase locking value (PLV) as the graph input. The GAT model outperforms state-of-the-art deep learning methods, demonstrating notable improvements with a two-class accuracy of 74.06% on an ALS dataset (approximately 320 trials collected over 1-2 months), and 71.89% on the BCI Comp IV 2a Dataset. This improvement demonstrates the effectiveness of graph-based representations to enhance classification performance for neurodegenerative conditions. There are statistically significant reductions in variance compared to state-of-the-art, due to subject-specific attention given by the model during testing. These results support the hypothesis that phase-locking value-based graph representations can enhance neural representations in BCIs, offering promising avenues for more personalized approaches in MI classification. This study highlights the potential for further optimizing GAT architectures and feature sets, pointing to future research directions that could improve performance and efficiency in MI classification tasks whilst establishing a lightweight methodology.