Roman Domination Based Graph Convolution Neural Network for Optimized EEG Signal Classification

This is the authors' abstract. We don't add key points for this paper.

An Electroencephalography (EEG) is a powerful, non-invasive tool for recording brain signals in Brain-Computer Interface (BCI) systems. However, handling multichannel data from numerous EEG signals can be challenging. One key complexity lies in identifying the channels most relevant to a specific task. Graph Convolution Networks (GCNs) have shown promising performance in classifying EEG-based complex tasks. However, current GCN-based approaches often rely on pre-defined graph adjacency matrices, which pose challenges in identifying highly correlated nodes and can lead to redundant information. Thereby increasing the computational time during training of the network. To address these limitations, our study introduces a novel approach by integrating Roman Domination a concept from graph theory into Graph Convolution Networks (RDGCN). This approach identifies key EEG channels that facilitate information flow, constructing an optimized graph structure that reflects the functional and topological relationships among channels. Furthermore, it reduces the computational load typically required to train a raw GCN. Unlike previous studies, we validated this innovative architecture on multiple divers datasets, including MODMA (with EEG data from subjects with and without depression) and BCI Competition IV 2a & 2b (featuring motor movement EEG recordings). Compared to other state-of-the-art techniques, proposed model demonstrated superior performance and also reduced 41.48% of computation time attributed to the application of Roman domination.