Exploring Non-Euclidean Approaches: A Comprehensive Survey on Graph-Based Techniques for EEG Signal Analysis

Abstract

Electroencephalogram (EEG) signals are widely applied in emotion recognition, sentiment analysis, disease classification, sleep disorder identification, and fatigue detection. Recent research has highlighted the active exploration of neurological disease analysis using EEG signals. Various machine learning and deep learning techniques, using feature-based and Euclidean approaches, have been employed to analyse these EEG signals. However, non-Euclidean approaches have proven more effective than Euclidean methods in EEG signal research. This superiority may stem from the nonlinear and dynamic characteristics of EEG signals, intricate interplay among brain regions, and resilience to common EEG signal noise. Unfortunately, limited studies on the graph representation of EEG signals exist due to constraints such as insufficient datasets, unavailable source code, and the complexity of graph representation. Hence, we aim to conduct a survey on various graph representation techniques, graph neural networks, existing methods, and available resources for EEG signal analysis using the non-Euclidean approach. In addition, visibility graph-based methods have been applied to singlechannel EEG signals, while graph neural networks have been shown to have promising outcomes in multichannel EEG signal analysis. Thus, the survey concluded that the non-Euclidean approach uses a graph to map more with the brain structure than with the Euclidean structure. Additionally, the inclusion of visibility graphs in multichannel EEG signals and graph neural networks would justify the robustness of the non-Euclidean approach in EEG signal analysis.