Multi-branch Domain Adversarial Neural Network with dynamic weight allocation for multi-source EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) data present significant challenges due to non-stationarity and inter-subject distribution shifts.
• Conventional Deep Domain Adaptation Network (DANN)-based methods struggle with feature representation and multi-source domain adaptation for EEG.
• Existing approaches lack robust solutions for cross-subject EEG classification, limiting their real-world applicability.

Purpose of the Study:

• To propose an advanced Multi-branch Domain Adversarial Neural Network with Multi-scale Channel Attention (MBCA-DANN) for EEG analysis.
• To enhance feature representation and improve multi-source domain adaptation capabilities in EEG classification.
• To address the limitations of current methods in handling non-stationarity and distribution shifts in EEG data.

Main Methods:

• Developed a Multi-scale Channel Attention Module (MSCA) to enrich feature representation and adaptively adjust feature channel weights.
• Constructed a multi-branch architecture incorporating auxiliary Maximum Mean Discrepancy (MMD), domain discriminators, and label discriminators for domain matching.
• Implemented a multi-source domain adaptation method with dynamic weight allocation to boost classification performance and robustness.

Main Results:

• Achieved 71.89% and 71.82% classification accuracy for single-source domain transfer on MII and MIII datasets, respectively.
• Improved multi-source domain transfer classification accuracy to 79.83% and 82.87% on the MII and MIII datasets.
• Attained 98.69% classification accuracy on a fatigue detection dataset, surpassing existing state-of-the-art algorithms.

Conclusions:

• The MBCA-DANN model demonstrates superior generalization ability for multi-source cross-subject EEG classification.
• The proposed method offers an effective solution for overcoming challenges in non-stationary and shifting EEG data distributions.
• MBCA-DANN significantly enhances classification performance and robustness compared to conventional DANN-based approaches.