Journal of Data Processing Ahmed Naddami, Hajar Ait Lamkademe 16 2 2026 https://doi.org/10.6025/jdp/2026/16/2/53-73 https://www.dline.info/jdp/fulltext/v16n2/jdpv16n2_1.pdf Brain computer interfaces (BCIs) have emerged as a transformative technology enabling direct communication between the human brain and external devices. Central to their effectiveness is the accurate decoding of electroencephalography (EEG) signals, which encapsulate complex neural dynamics across spatial and temporal scales. However, EEG signals are inherently noisy, high-dimensional, non-stationary, and characterized by irregular spatial structures, making their analysis particularly challenging. Traditional deep learning approaches, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have demonstrated considerable success in EEG classification tasks. Nevertheless, these models operate under Euclidean assumptions and fail to adequately capture the intrinsic non-Euclidean topology of brain connectivity. In contrast, Graph Neural Networks (GNNs) provide a principled framework for modeling such data by explicitly incorporating relationships between EEG electrodes through graph structures. In this work, EEG signals are modeled as graphs, where electrodes correspond to nodes and functional relationships define edges. A comprehensive formulation of GNN-based learning is presented, including spectral and spatial graph convolutions, attention mechanisms, and spatiotemporal extensions. Furthermore, embedding space analysis is discussed to enhance the interpretability of learned representations. The paper synthesizes recent advancements in graph-based EEG modeling while highlighting key challenges and future research directions.