APCformer: an aggregation-perception enhanced convolutional transformer network for MI-EEG decoding.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) decoding is crucial for Brain-Computer Interfaces (BCI).
• Existing methods struggle with spatial-temporal feature extraction and balancing long-range/local dependencies in EEG signals.
• This leads to information loss and difficulty adapting to EEG's temporal dynamics.

Purpose of the Study:

• To propose a novel network, the Aggregation-Perception Enhanced Convolutional Transformer (APCformer), to address limitations in current EEG decoding methods.
• To enhance the extraction of fine-grained spatial-temporal features and improve the modeling of both global and local EEG signal characteristics.
• To boost the accuracy and efficiency of EEG decoding for BCI applications.

Main Methods:

• Developed the APCformer network featuring a branch-interactive structure for joint shallow feature extraction using multi-scale spatial-temporal convolution.
• Integrated an Adaptive Feature Recalibration (AFR) module for cross-scale feature interaction and fine-grained feature enhancement.
• Employed a Position-aware Enhancement (PAE) module with learnable positional encoding to improve temporal relationship characterization and a Sparse Information Aggregation Transformer (SAT) for balanced feature modeling.

Main Results:

• APCformer achieved superior performance on public BCI-IV 2a and BCI-IV 2b datasets.
• Average decoding accuracies reached 85.53% on BCI-IV 2a and 89.15% on BCI-IV 2b.
• Demonstrated strong capability in handling complex EEG features and dynamic patterns.

Conclusions:

• The proposed APCformer network effectively overcomes limitations of existing EEG decoding methods.
• APCformer significantly improves the accuracy and efficiency of EEG decoding tasks.
• The network's architecture enhances the capture of crucial spatial-temporal dynamics and temporal characteristics in EEG signals.