Geo-GCN: Geometric-Graphical Convolutional Network for EEG-based Auditory Attention Detection.

Area of Science:

• Neuroscience
• Machine Learning
• Bioengineering

Background:

• Auditory attention detection (AAD) uses electroencephalography (EEG) to infer speech focus.
• Current methods often lack biological plausibility in feature extraction.

Purpose of the Study:

• To introduce a novel geometric graph convolutional network (Geo-GCN) for enhanced EEG-based AAD.
• To leverage the physical sensor layout for more biologically informed feature learning.

Main Methods:

• Developed Geo-GCN utilizing sensor distances for adjacency matrix construction.
• Applied Geo-GCN to EEG data from normal hearing (NH) and hearing-impaired (HI) participants.
• Compared Geo-GCN performance against standard graph convolutional networks (GCNs).

Main Results:

• Geo-GCN significantly outperformed traditional GCNs in AAD accuracy.
• The proposed method demonstrated reduced performance variability across participants.
• Consistent performance gains were observed in both NH and HI groups.

Conclusions:

• Explicitly modeling scalp geometry with Geo-GCN improves EEG-based auditory attention detection.
• Geometry-aware graph neural networks show promise for heterogeneous populations, including those with hearing impairments.
• This approach offers a more robust and biologically informed solution for AAD.