EEG-based Auditory Attention Switch Detection with Multi-scale Gated Attention and Multi-task Learning based...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Auditory attention switch detection (AASD) is crucial for adaptive neurotechnologies but faces challenges with EEG's low signal-to-noise ratio (SNR).
• Existing methods struggle with feature discriminability and high detection delays, limiting their effectiveness.

Purpose of the Study:

• To propose a novel Hierarchical Spatiotemporal Network (HSTN) for enhanced auditory attention switch detection from EEG signals.
• To improve feature extraction and reduce detection delay in complex auditory environments.

Main Methods:

• Developed a Hierarchical Spatiotemporal Network (HSTN) utilizing a spatiotemporal encoder and multi-scale gated attention.
• Implemented a multi-task joint training strategy for synchronous optimization of auditory attention switch detection and decoding.
• Extracted spatiotemporal features integrating short-term and long-term dependencies.

Main Results:

• HSTN significantly outperformed baseline models in auditory attention switch detection (F1=0.89, accuracy 88.6%) and auditory attention decoding (accuracy 89.3%).
• Demonstrated superior model parameter efficiency and reduced inference time compared to existing methods.
• Ablation studies confirmed the importance of multi-task learning, gated attention, and multi-scale convolutions.

Conclusions:

• The proposed HSTN offers an efficient and effective solution for EEG-based auditory attention switch detection in complex scenarios.
• Spatiotemporal feature encoding combined with multi-task learning provides a practical framework for intelligent hearing aids and auditory brain-computer interfaces.