EEG-Based Motor Imagery Recognition Framework via Multisubject Dynamic Transfer and Iterative Self-Training.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) models often require subject-specific calibration, limiting their practical application, especially in motor imagery (MI) tasks.
• Existing electroencephalogram (EEG) decoding models struggle with variations in subjects and time periods, necessitating extensive data collection and retraining.

Purpose of the Study:

• To develop an unsupervised domain adaptation framework, Iterative Self-Training Multi-Subject Domain Adaptation (ISMDA), for robust offline motor imagery (MI) decoding.
• To overcome the limitations of subject and period variations in EEG decoding for BCI applications.

Main Methods:

• Designed a feature extractor to map EEG data into a discriminative latent space.
• Employed a dynamic transfer attention module for improved source and target domain sample matching.
• Utilized iterative self-training with a target-domain classifier and a confidence-based pseudolabeling algorithm.

Main Results:

• Achieved high cross-subject classification accuracies: 69.51% (BCI IV IIa), 82.38% (High gamma), and 90.98% (Kwon et al.).
• Demonstrated superior performance compared to current state-of-the-art offline algorithms for MI tasks.
• Successfully addressed the challenges posed by subject and period variations in offline MI BCI.

Conclusions:

• The proposed ISMDA framework offers a robust solution for unsupervised domain adaptation in BCI.
• This method significantly improves the efficiency and applicability of EEG-based MI decoding, particularly for rehabilitation.
• The findings pave the way for more practical and adaptable BCI systems.