Frequency-Aware Spatial-Temporal Attention Explainable Network for EEG Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Representation learning in spatial and temporal domains has advanced electroencephalography (EEG) decoding for brain-computer interfaces (BCIs).
• The significance of frequency information, vital to neurological mechanisms, has been largely underutilized in prior EEG decoding models.
• Existing methods often neglect the synergistic integration of frequency, spatial, and temporal data inherent in EEG signals.

Purpose of the Study:

• To propose FSTNet, a novel neural network architecture that synergistically integrates frequency, spatial, and temporal domains for enhanced EEG decoding.
• To adaptively learn informative frequency signatures from broadband EEG signals, emphasizing the importance of frequency information.
• To improve the accuracy and transparency of EEG decoding by capturing discriminative neural signatures.

Main Methods:

• FSTNet utilizes broadband EEG signals as input, adaptively learning informative frequency signatures.
• A frequency-aware module assigns selective weights to latent representations in the frequency space, highlighting frequency information.
• Self-attention mechanisms are employed to capture spatial and temporal dependencies, extracting discriminative features for decoding.

Main Results:

• FSTNet achieved superior results on motor imagery and emotion recognition tasks across SEED, PhysioNet, and OpenBMI datasets.
• The model demonstrated strong performance in both individual and cross-subject EEG decoding scenarios.
• Visualizations confirmed that FSTNet captures task-specific frequency ranges and spatial patterns, aligning with physiological mechanisms.

Conclusions:

• The proposed FSTNet effectively decodes EEG signals by synergistically leveraging frequency, spatial, and temporal information.
• The method enhances the transparency of the learning process by visualizing captured frequency and spatial patterns.
• FSTNet shows significant potential for advancing EEG decoding and deepening the understanding of neurological processes.