Brain connectivity and time-frequency fusion-based auditory spatial attention detection.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory spatial attention detection (ASAD) uses electroencephalogram (EEG) signals to identify auditory attention locus.
• Current ASAD models struggle with EEG feature extraction, leading to overfitting and reduced discriminability.
• Existing models often overlook topological relationships and brain connectivities between EEG channels.

Purpose of the Study:

• To develop an advanced ASAD model that overcomes limitations in EEG feature extraction and connectivity analysis.
• To improve the precision and discriminability of EEG representations for ASAD.
• To effectively integrate local and global brain connectivities into graph-based EEG modeling for ASAD.

Main Methods:

• Proposed a novel ASAD model incorporating time-frequency feature fusion for enhanced EEG representation.
• Treated EEG segments as graphs, utilizing graph convolution and global attention mechanisms to capture local and global brain connections.
• Conducted experiments using leave-trials-out cross-validation on MAD-EEG, KUL, and SNHL datasets.

Main Results:

• Achieved >9% and >3% higher accuracy than state-of-the-art models on MAD-EEG and KUL datasets, respectively.
• Demonstrated comparable accuracy to state-of-the-art models on the SNHL dataset.
• Highlighted the indispensable role of EEG time-frequency feature fusion and identified frontal and temporal lobe electrodes as crucial for ASAD.

Conclusions:

• The proposed ASAD model offers superior performance, particularly on smaller datasets, by effectively integrating EEG time-frequency features and brain connectivity information.
• The findings advance the understanding of neural encoding in human hearing and attention.
• Potential applications include the development of neuro-steered hearing devices.