GCANet: Enhancing EEG-based auditory attention decoding with temporal frequency GCN and cross attention mechanisms

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Selective auditory attention, or the cocktail party effect, enables focus on target speakers amid noise.
• Auditory attention decoding (AAD) aims to identify attended speakers from electroencephalography (EEG) signals.
• Existing AAD methods often neglect the graph structure inherent in EEG data.

Purpose of the Study:

• To propose GCANet, an end-to-end model that leverages graph convolutional networks and cross-attention for improved AAD.
• To capture functional brain connectivity and enhance EEG-audio feature interactions.
• To evaluate GCANet's performance on public datasets for cross-trial and cross-subject decoding.

Main Methods:

• Developed GCANet, integrating a time-frequency graph convolutional network (TFGCN) for brain connectivity analysis.
• Incorporated a cross-attention mechanism to dynamically fuse EEG and audio features.
• Validated the model on KUL, DTU, and AVGC datasets using 1-second decision windows.

Main Results:

• GCANet achieved high decoding accuracies: 92.2% (KUL), 83.2% (DTU), and 62.6% (AVGC) in cross-trial settings.
• Cross-subject accuracies reached 75.1% (KUL), 57.1% (DTU), and 55.6% (AVGC).
• Analysis revealed potential gaze-related confounds and highlighted frontal and temporal regions in EEG-audio interactions.

Conclusions:

• GCANet significantly enhances auditory attention decoding accuracy by modeling brain connectivity and cross-modal interactions.
• Findings suggest potential confounds in AAD related to visual cues and identify key brain regions involved.
• The study provides valuable insights into cross-modal EEG-audio interactions and future AAD research.