EEG-Deformer: A Dense Convolutional Transformer for Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Computer Science
• Signal Processing

Background:

• Learning temporal dynamics in electroencephalogram (EEG) signals is crucial for brain-computer interfaces (BCIs).
• Existing Transformer-CNN models struggle to capture the essential coarse-to-fine temporal dynamics of EEG signals.
• There is a need for advanced methods to improve the decoding accuracy of brain activities.

Purpose of the Study:

• To introduce EEG-Deformer, a novel CNN-Transformer architecture designed to effectively learn temporal dynamics in EEG signals.
• To overcome the limitations of current methods in capturing coarse-to-fine temporal patterns.
• To enhance the decoding accuracy of brain activities for various cognitive tasks.

Main Methods:

• The proposed EEG-Deformer incorporates a Hierarchical Coarse-to-Fine Transformer (HCT) block with a Fine-grained Temporal Learning (FTL) branch.
• A Dense Information Purification (DIP) module is utilized to enhance decoding accuracy using multi-level, purified temporal information.
• The architecture integrates these novel components into a CNN-Transformer framework.

Main Results:

• EEG-Deformer demonstrated superior or comparable performance to state-of-the-art methods across three cognitive tasks: cognitive attention, driving fatigue, and mental workload detection.
• Comprehensive experiments confirmed the generalizability of the proposed EEG-Deformer.
• Visualization results indicated that EEG-Deformer learns from neurophysiologically meaningful brain regions relevant to the specific cognitive tasks.

Conclusions:

• EEG-Deformer effectively captures coarse-to-fine temporal dynamics in EEG signals, outperforming existing methods.
• The novel HCT block and DIP module contribute to enhanced decoding accuracy in BCIs.
• The model's ability to learn from meaningful brain regions highlights its potential for advanced BCI applications.