[Paper Review] Graph Convolutional Neural Networks for analysis of EEG signals, BCI application

Mirfarid Musavian Ghazani, Anh Huy Phan|arXiv (Cornell University)|Jun 16, 2020

EEG and Brain-Computer Interfaces37 references4 citations

TL;DR

This paper proposes DeepCSP, a differentiable, non-linear extension of Common Spatial Patterns (CSP) as a novel loss function for end-to-end training of Graph Convolutional Neural Networks (GCNs) on raw EEG signals for Motor Imagery BCI applications. By learning linearly separable spatial filters in a differentiable manner, DeepCSP improves classification accuracy on EEG datasets, with GCN-based models achieving state-of-the-art performance on benchmark datasets like BCI Competition IV-2a and Motor Imagery dataset.

ABSTRACT

Decoding brain signals has gained many attention and has found much applications in recent years such as Brain Computer Interfaces, communicating with controlling external devices using the user's intentions, occupies an emerging field with the potential of changing the world, with diverse applications from rehabilitation to human augmentation. This being said brain signal analysis, EEG brain signal analysis in particular, is a challenging task. With the advances and achievements in the field of deep learning in problem solving with using only raw data, few attempts has been carried in recent years, to apply deep learning to tackle EEG among other types of brain signals. In this study, we propose a novel loss function, called DeepCSP to extend the classical Common Spatial Patterns to a non linear, differentiable module to serve as the loss function to enforce linearly separable latent representations of EEG signals belonging to different classes in an end to end manner on raw signals without the need to perform extensive feature engineering. With recent generalizations of deep learning methods to work on arbitrarily structured graphs and the introduced loss we have proposed two light weight models to decode EEG signals and carried experiments to show their performance.

Motivation & Objective

• To address the challenge of manual feature engineering in EEG signal classification for BCI applications.
• To develop a differentiable, non-linear extension of the classical CSP method to enable end-to-end training with deep neural networks.
• To improve classification performance of EEG signals by integrating graph convolutional networks with a novel loss function.
• To explore the effectiveness of representing EEG signals as graphs for improved feature learning in motor imagery tasks.
• To enable subject-independent BCI applications through scalable, learnable spatial filtering.

Proposed method

• Proposes DeepCSP, a differentiable loss function that generalizes classical CSP to a non-linear, learnable module for end-to-end training.
• Represents EEG signals as graphs using connectivity matrices derived from phase-locking value (PLV) to model inter-electrode relationships.
• Employs GraphSAGE as the GCN backbone to extract spatial-temporal features from graph-structured EEG data.
• Trains a lightweight CNN-GCN architecture with separate learning rates: 0.01 for CNN/GCN and 0.1 for the classifier head.
• Uses the spatial filters learned via DeepCSP to generate topographic maps (topomaps) for interpreting brain activation patterns.
• Employs a two-stage training process: first, spatial filters are optimized via DeepCSP on the full training set to improve generalization.

Experimental results

Research questions

• RQ1 Can a differentiable, non-linear extension of CSP improve end-to-end learning in EEG-based BCI systems?
• RQ2 How does representing EEG signals as graphs enhance classification performance in motor imagery tasks?
• RQ3 What is the impact of different connectivity estimation methods (e.g., PLV) on GCN-based EEG classification?
• RQ4 Can DeepCSP learn spatial filters that align with known neurophysiological patterns (e.g., motor cortex activation)?
• RQ5 How does the proposed method generalize across subjects, especially in subject-independent settings?

Key findings

• The proposed DeepCSP loss significantly improves classification accuracy on the BCI Competition IV-2a dataset compared to baseline methods.
• On the Motor Imagery dataset, the model with DeepCSP and GCN-based feature extraction achieved higher accuracy than standard CSP and traditional methods.
• Using PLV for connectivity matrix estimation yielded better performance than other methods, confirming its suitability for EEG graph modeling.
• The spatial filters learned via DeepCSP produced interpretable topographic maps showing clear activation patterns over C3 and C4 electrodes for right and left hand motor imagery.
• The model with 4 components in the spatial filter layer achieved optimal performance, indicating a balance between complexity and generalization.
• Dynamic graph convolutional networks were found to overfit quickly due to high parameter count and noisy EEG data, leading to poor performance.