Source-linked AI summary
Graph Neural Network-based EEG Classification: A Survey
Dominik Klepl, Min Wu, Fei He
TL;DR
EEG-GNN research spans diverse designs, creating a need for a systematic review of the methods and their trends. The paper categorizes these approaches and identifies constraints on generalisability to external datasets with different numbers of EEG signals.
Problem
A growing range of GNN methods for EEG classification calls for a systematic categorization of their designs and trends.
Source: §I · p. 1, §IX · p. 9Method
The review categorizes EEG-GNN models by brain graph structure, node features and preprocessing, graph convolution, pooling, and graph-embedding formation.
Source: §IX · p. 9Results
The survey finds limited variety in graph-convolution layers, with many papers using ChebConv or vanilla spatial GCN.
Source: A. Limitations of Surveyed Papers · p. 9Takeaways & Limitations
The review identifies transfer learning and handling varying numbers of EEG signals or electrodes as directions for improving model generalisability.
Source: A. Limitations of Surveyed Papers · p. 9Many surveyed models may not generalise to external datasets with different numbers of EEG signals because they use ChebConv and node-feature concatenation.
Source: A. Limitations of Surveyed Papers · p. 9
Abstract
from arXiv · showhide
Graph neural networks (GNN) are increasingly used to classify EEG for tasks such as emotion recognition, motor imagery and neurological diseases and disorders. A wide range of methods have been proposed to design GNN-based classifiers. Therefore, there is a need for a systematic review and categorisation of these approaches. We exhaustively search the published literature on this topic and derive several categories for comparison. These categories highlight the similarities and differences among the methods. The results suggest a prevalence of spectral graph convolutional layers over spatial. Additionally, we identify standard forms of node features, with the most popular being the raw EEG signal and differential entropy. Our results summarise the emerging trends in GNN-based approaches for EEG classification. Finally, we discuss several promising research directions, such as exploring the potential of transfer learning methods and appropriate modelling of cross-frequency interactions.
I. INTRODUCTION
EEG classification is challenged by noisy, nonstationary, multivariate signals, while manually engineered features and conventional deep-learning approaches have limitations. This review surveys GNN methods for EEG classification, including graph convolutional layers and emerging design trends.
I. INTRODUCTION
- EEG signals pose low signal-to-noise, nonstationarity, and multivariate challenges across classification tasks such as emotion recognition, motor imagery, and neurological disorders.Source: §I · p. 1
- Manually defined EEG features can introduce selection bias and time-consuming engineering, limiting scalability and generalisation across datasets.Examples include common spatial patterns, wavelet and Hilbert-Huang transforms, and power spectral density.Source: §I · p. 1
- CNNs and LSTMs can struggle to capture spatial dependencies between electrodes and EEG temporal dynamics, both relevant to classification.The text describes modelling sequential and spatial relations as crucial for more accurate classification and analysis.Source: §I · p. 1
- GNNs are presented as a graph-structured approach that can leverage electrode relations to extract features and make predictions from EEG.The review frames this approach within network neuroscience and EEG graph modelling.Source: §I · p. 1
- The review aims to address growing EEG-GNN literature by identifying trends and reviewing graph-convolution layers and their applicability to EEG.Its stated contributions also include a unified overview of node features and brain graph definitions, graph-embedding formation, and future research directions.Source: §I · pp. 1–2
II. OVERVIEW OF GRAPH NEURAL NETWORKS
The paper frames EEG classification as graph classification, with EEG channels as nodes and connectivity as edges. It outlines spatial GNN message passing, whose receptive field expands with depth but can be constrained by oversmoothing.
II. OVERVIEW OF GRAPH NEURAL NETWORKS
- EEG graph classification treats channels as nodes, structural or functional connectivity as edges, node features as input, and the class label as the prediction target.The paper represents a featured graph as G = (V, E, H) and the classification function as f(G) → y.Source: §II · p. 2
- GNNs accept both node features and graph structure and can accommodate irregular graphs with varying numbers of nodes, unlike fixed-size-input CNNs.Source: §II · p. 2
A. Spatial GNNs
- Spatial GNNs update node embeddings by aggregating neighbourhood information through message passing, with each layer accessing one-hop neighbours.Stacking L layers extends access to up to L-hop neighbourhoods, but deep models can oversmooth as node embeddings become increasingly similar.Source: A. Spatial GNNs · p. 3
B. Spectral GNNs
Spectral GNNs transform graph signals into a frequency representation and learn filters over graph frequencies. ChebConv offers a localized approximation, while the paper notes spectral GNNs' fixed-node-count constraint.
B. Spectral GNNs
- Spectral GNNs transform EEG graph features into the graph spectral domain using the Graph Fourier Transform and graph signal processing.The graph spectrum is derived from the Laplacian eigendecomposition; the inverse transform maps the filtered signals back to the graph domain.Source: B. Spectral GNNs · p. 3
- Learnable spectral filters weight graph-frequency contributions before the filtered signals are transformed back to node embeddings.The paper describes spectral convolution as element-wise multiplication in the spectral domain.Source: B. Spectral GNNs · p. 3, Figure 2 · p. 4
- ChebConv approximates full spectral convolution with localized filtering, while spectral GNNs are limited to input graphs with a fixed number of nodes.The Chebyshev parameter K controls filter size; full spectral GNNs can access N-hop information, whereas approximations restrict access to the chosen hop size.Source: B. Spectral GNNs · p. 3, Figure 2 · p. 4
III. SURVEY RESULTS
The survey reviews 63 articles on EEG-GNN classification and categorises methods by their model modules. It focuses on those modules rather than task-specific architecture, since classification task is considered to have a relatively minor effect on design.
III. SURVEY RESULTS
- The review covers 63 articles selected by title and abstract screening from Google Scholar and ScienceDirect, including peer-reviewed articles and preprints across EEG classification tasks.The search was run on November 1, 2022, using terms for graph neural or convolutional networks and EEG.Source: §III · p. 4
- Emotion recognition, epilepsy diagnosis and detection, and motor imagery are the most common surveyed classification tasks.The survey treats task type as having a relatively minor effect on GNN architecture and therefore does not analyse tasks in detail.Source: §III · p. 4
- The survey categorises EEG-GNN methods by brain graph structure, node features, graph convolutional layers, preprocessing, pooling, and graph embedding formation.These categories are intended to support comparison and identify emerging trends and future directions.Source: §III · p. 5
IV. DEFINITION OF BRAIN GRAPH STRUCTURE
The survey categorises brain graph definitions as learnable or pre-defined, and also distinguishes structural from functional connectivity. It highlights trade-offs in noise sensitivity, graph coverage, and compatibility with varying input sizes.
IV. DEFINITION OF BRAIN GRAPH STRUCTURE
- Brain graph definitions are categorised as learnable or pre-defined, with structural connectivity generally pre-defined and functional connectivity either pre-defined or learnable.Here, structural connectivity refers to electrode physical distances, while functional connectivity refers to pairwise statistical relationships between EEG signals.Source: §IV · p. 5
- Distance-based structural graphs use electrode distance and a manually set threshold to control graph sparsity.The supplied text identifies the edge weight, electrode distance, and threshold, but does not specify the full edge-weight formula.Source: §IV · p. 5
- Structural graphs are signal-independent and consistent across samples using the same montage, but their short-range edges may not represent the underlying brain network accurately.Some surveyed papers add global or inter-hemispheric edges to address this limitation.Source: §IV · p. 5
- Functional-connectivity graphs are sample-specific, can contain short- and long-range edges, and may be sensitive to noise because they derive from EEG signals.The survey includes both classical connectivity measures and learnable methods among functional-connectivity approaches.Source: §IV · p. 5
- A shared learnable mask requires a fixed number of input nodes because it must match the input adjacency matrix’s shape.The survey says there is no clear preferred graph-definition method; researchers should consider each method’s strengths and weaknesses for their task.Source: §IV · pp. 5–6
V. NODE FEATURE DEFINITIONS
The survey groups EEG node features by time, frequency, and graph domains. Time-domain differential entropy and raw EEG signals are most common, while the alternatives have distinct dimensionality or graph-definition constraints.
V. NODE FEATURE DEFINITIONS
- Time-domain differential entropy and raw EEG signals are the most commonly used node features in the surveyed literature.The popularity of differential entropy is linked to its inclusion in open EEG datasets such as SEED.Source: §V · p. 6
- Raw EEG signals can be too long for effective GNN processing, so papers often pair them with preprocessing or spatio-temporal GNNs to reduce dimensionality or extract temporal patterns.Descriptive statistics such as mean, median, and standard deviation are alternatives.Source: §V · p. 6
- Fourier frequency components and power spectral density quantify EEG frequency-component strength and have relatively low dimensionality compared with raw signals.Source: §V · p. 6
- Graph-theoretical node features include mean node weight and betweenness centrality, but require a pre-defined graph and are incompatible with learnable brain-graph methods.Source: §V · pp. 6–7
A. Node Feature Preprocessing
Node feature preprocessing is optional and is usually integrated into the GNN architecture. The surveyed preprocessing modules use neural networks to reduce or enhance node features, with pre-trained feature extractors as exceptions.
A. Node Feature Preprocessing
- Node feature preprocessing is an optional step before GNN layers, and the survey summarises its types in Table III.Source: A. Node Feature Preprocessing · p. 7
- Most preprocessing modules are integrated into the GNN for end-to-end training; exceptions use pre-trained bidirectional LSTM or CNN feature extractors.Source: A. Node Feature Preprocessing · p. 7
- The surveyed preprocessing modules use neural networks, usually CNN variants or MLPs, to reduce dimensionality or enhance node features, potentially suppressing noise or redundant information.Source: A. Node Feature Preprocessing · p. 7
VI. TYPE OF GRAPH CONVOLUTIONAL LAYER
The survey groups graph-convolution approaches for EEG classification into spatial, spectral, and spatio-temporal forms, with some studies also using parallel multi-branch architectures. ChebConv is used by a majority of surveyed papers, while the spatial-GNN papers explore a wider range of methods.
VI. TYPE OF GRAPH CONVOLUTIONAL LAYER
- The survey categorizes GCN layers as spatial or spectral, with temporal layers treated as an addition that must combine with one of those types.Table IV summarizes the GCN types used in the surveyed papers.Source: §VI · p. 7
- ChebConv appears in the majority of surveyed papers, although the survey finds it unclear why authors choose this spectral-GNN approximation.The authors note that EEG graphs are relatively small, so a full spectral GNN would not be too computationally expensive; they speculate that classical signal-processing influences may explain the choice.Source: §VI · p. 7
- The other half of surveyed papers experiment with spatial GNNs, including simplified GCN and graph attention networks that reweight edges.A simplified GCN is equivalent to a first-order ChebConv (K = 1).Source: §VI · p. 7
- Spatio-temporal blocks combine one graph-convolution layer with a temporal 1D-CNN to extract spatial and temporal patterns.Both spatial and spectral variants are used, but the survey reports no comparative study establishing which is preferable.Source: §VI · p. 7
- Multi-branch architectures apply multiple GCN layers in parallel, including models using functional- and structural-connectivity graphs or frequency-band views.The cited examples use two branches for FC- and SC-based structure, or separate graph views constructed from EEG frequency bands.Source: §VI · p. 7
VII. NODE POOLING MECHANISMS
The survey describes node pooling as a relatively unexplored way to reduce graph size or identify important nodes. For graph classification, it also reviews ways to turn learned node embeddings into a unified graph representation, including concatenation and permutation-invariant readouts.
VII. NODE POOLING MECHANISMS
- Node pooling can remove redundant nodes, reduce concatenation-based graph embeddings, and identify node importance for explainability.The surveyed modules include both learnable and non-learnable approaches.Source: §VII · pp. 7–8
- Node pooling remains relatively unexplored in EEG-GNN classification models.The survey summarizes the pooling modules used in the papers in Table V.Source: §VII · p. 7
VIII. FROM NODE EMBEDDINGS TO GRAPH EMBEDDING
- Graph classification requires transforming the learned node embeddings into a unified graph representation, unlike node classification or link prediction.The survey summarizes these transformation methods in Table VI.Source: §VIII · p. 8
- Concatenating node features makes the graph embedding grow with node count, requires equal node counts across input graphs, and can retain redundant information.The survey links the redundancy to GNN node embeddings aggregating information from neighboring nodes.Source: §VIII · p. 8
- A readout forms graph embeddings by applying a permutation-invariant function to the set of node embeddings.Surveyed readouts include sum, average, maximum, and attention-weighted sum; other approaches use neural networks or graph pooling methods.Source: §VIII · p. 8
IX. DISCUSSION
The discussion surveys EEG-GNN applications and organizes methods by their inputs and modules. It also identifies limitations and research gaps, including limited GCN-layer variety, generalization concerns, a lack of transfer-learning experiments, and underexplored cross-frequency modeling.
IX. DISCUSSION
- GNNs have been applied to diverse EEG classification tasks, and can model multivariate time series and their interactions while producing patterns interpretable through network neuroscience.The survey mentions emotion recognition, brain-computer interfaces, and psychological and neurodegenerative disorders and diseases.Source: §IX · pp. 8–9
- The survey categorizes methods by brain graph structure, node features and preprocessing, GCN layers, node pooling, and graph-embedding formation.The authors say this organization provides an overview of the literature and helps identify future directions.Source: §IX · p. 9
A. Limitations of Surveyed Papers
- The survey finds the least variety and innovation in GCN layers, with many papers using ChebConv or vanilla spatial GCN.The authors suggest the field’s relative novelty may contribute; a few papers test more complex layers and multi-branch architectures.Source: A. Limitations of Surveyed Papers · p. 9
- Most surveyed papers have limited generalizability to external datasets with different numbers of EEG signals, associated with ChebConv and node-feature concatenation.The survey suggests spatial GCN layers, readout functions, or pooling to a fixed node count as ways to address these constraints.Source: A. Limitations of Surveyed Papers · p. 9
- The survey identifies a general lack of transfer-learning experiments in EEG-GNN models as a possible future research direction.The authors describe transfer learning as promising, not as an established result.Source: A. Limitations of Surveyed Papers · p. 9
- The surveyed work rarely models frequency-band information in more complex ways, with examples limited to separate band-specific models or concatenated branch embeddings.The cited approaches train separate models for frequency bands or concatenate embeddings from frequency-band GNN branches.Source: A. Limitations of Surveyed Papers · p. 9
B. Future Directions
The survey proposes future work on comparing graph-convolution layers, improving generalisability through transfer learning, and incorporating cross-frequency information. It also notes that its overview is not exhaustive and omits comprehensive treatment of model explainability.
B. Future Directions
- Systematically compare spatial GNN, ChebConv, GAT, and graph-transformer layers to clarify how layer choice influences classification performance.The survey identifies this as a crucial design question for future work.Source: B. Future Directions · p. 9
- Transfer learning and handling varying numbers of EEG signals or electrodes are proposed to improve generalisability across datasets and tasks.The authors suggest pretraining on large, accessible emotion-recognition or BCI datasets for data-limited clinical and rare-disease applications.Source: B. Future Directions · p. 9
- Integrating cross-frequency coupling into EEG-GNN models is proposed as a way to use EEG frequency information more richly.The survey links neuroscience findings on CFC with potential performance and explainability gains.Source: B. Future Directions · pp. 9–10
C. Limitations of Our Survey
- Because the survey does not use systematic-review methodology, its findings are not asserted to be exhaustive.Source: C. Limitations of Our Survey · p. 10
- The survey omits comprehensive coverage of model explainability and suggests a comparative experimental study for examining explainability options.It lists papers reporting use of certain explainability methods:.Source: C. Limitations of Our Survey · p. 10
X. CONCLUSION
- The paper concludes by examining current research on EEG-GNN models for EEG classification.Source: §X · p. 10