Unsupervised Domain Adaptation With Synchronized Self-Training for Cross- Domain Motor Imagery Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Robust brain-computer interface (BCI) systems require effective electroencephalography (EEG) decoding.
• Current EEG models struggle with data heterogeneity across different experimental domains.
• This limits BCI adaptability in real-world, variable conditions.

Purpose of the Study:

• To introduce Synchronized Self-Training Domain Adaptation (SSTDA) for cross-domain motor imagery classification.
• To enhance EEG decoding performance by addressing data distribution discrepancies.
• To develop a method for robust BCI deployment across diverse scenarios.

Main Methods:

• SSTDA utilizes labeled source domain data and unlabeled target domain data via self-training.
• A feature extractor maps raw EEG signals to a latent space for representation learning.
• An easy-to-hard self-training process optimizes the feature extractor for a domain-shared latent space.

Main Results:

• SSTDA achieved high classification accuracies in inter-subject tasks (64.43% and 80.40%) on public datasets.
• The method outperformed existing approaches in inter-session and inter-dataset validation.
• Tested on a new six-class motor imagery dataset, SSTDA reached accuracies of 77.09% and 80.18%.

Conclusions:

• SSTDA demonstrates superior performance in cross-domain EEG decoding compared to existing algorithms.
• The approach effectively learns discriminative, domain-invariant representations.
• SSTDA significantly enhances BCI system robustness and generalizability for motor imagery tasks.