Lightweight Source-Free Domain Adaptation Based on Adaptive Euclidean Alignment for Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) require robust cross-subject recognition for privacy protection.
• Source-free domain adaptation (SFDA) is effective but updating models for each new subject is inconvenient.
• Domain drift between subjects hinders the performance of subject-independent BCIs.

Purpose of the Study:

• To propose Adaptive Euclidean Alignment (AEA) to address domain drift in SFDA for cross-subject EEG classification.
• To integrate AEA with existing SFDA methods to create novel, improved BCI models.
• To evaluate the efficacy of AEA-based SFDA methods across diverse EEG datasets and deep learning architectures.

Main Methods:

• Extended Euclidean Alignment (EA) to propose Adaptive Euclidean Alignment (AEA), learning a projection matrix to align target and source subject data distributions.
• Combined AEA with SHOT, GSFDA, and NRC to develop AEA-SHOT, AEA-GSFDA, and AEA-NRC.
• Applied these AEA-based SFDA methods to EEGNet, Shallow ConvNet, Deep ConvNet, and MSFBCNN on motor imagery, ERP, and SSVEP datasets.

Main Results:

• AEA effectively eliminates domain drift, significantly improving cross-subject EEG classification performance.
• AEA-based SFDA methods demonstrated superior performance compared to existing approaches.
• AEA-SHOT achieved the highest average accuracy of 81.4% on the PhysioNet dataset.

Conclusions:

• The proposed AEA is a powerful technique for enhancing subject-independent BCI performance by mitigating domain shift.
• AEA-based SFDA methods offer a practical and effective solution for cross-subject EEG recognition, reducing computational burden.
• These findings highlight the potential of AEA for advancing privacy-preserving and high-performance BCIs.