Dynamic decomposition graph convolutional neural network for SSVEP-based brain-computer interface.

Area of Science:

• Neuroscience
• Computer Science
• Signal Processing

Background:

• Steady-state visually evoked potential (SSVEP) paradigms are common in brain-computer interfaces (BCI).
• Processing multi-channel electroencephalogram (EEG) data presents challenges due to its non-Euclidean nature, requiring methods that consider inter-channel topology.
• Existing methods often struggle with issues like oversmoothing and gradient vanishing in graph convolutional networks (GCNs).

Purpose of the Study:

• To introduce a novel Dynamic Decomposition Graph Convolutional Neural Network (DDGCNN) for classifying SSVEP EEG signals.
• To address challenges in EEG data processing, including non-Euclidean characteristics and inter-channel topological relations.
• To improve feature extraction and adaptive aggregation in SSVEP-based BCIs.

Main Methods:

• Developed DDGCNN incorporating layerwise dynamic graphs to combat oversmoothing in GCNs.
• Employed a dense connection mechanism to mitigate gradient vanishing.
• Integrated graph dynamic fusion to enhance traditional linear transformations for improved feature extraction and aggregation.
• Processed SSVEP time-domain signals directly in an end-to-end system.

Main Results:

• DDGCNN demonstrated superior performance in learning and extracting features from EEG topological structures.
• The proposed DDGCNN outperformed state-of-the-art (SOTA) algorithms on two benchmark datasets.
• Successfully showcased DDGCNN implementation for synchronized BCI robotic fish control.

Conclusions:

• DDGCNN represents a significant advancement in EEG signal processing for SSVEP-based BCIs.
• The end-to-end, direct processing of SSVEP time-domain signals makes DDGCNN easily deployable.
• The method effectively handles the non-Euclidean nature of EEG data and improves classification accuracy.