DA-META: A Dual Attention Meta-Learning Framework for Unsupervised Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery electroencephalography (MI-EEG) decoding shows promise for paralysis rehabilitation.
• Generalization challenges in MI-EEG are caused by intersubject variability and limited labeled data.
• Existing meta-learning approaches for unsupervised domain adaptation have limitations in feature extraction and target data utilization.

Purpose of the Study:

• To propose a novel dual-attention meta-learning framework (DA-META) to address limitations in MI-EEG decoding.
• To improve the generalization capability of MI-EEG decoding models in unsupervised domain adaptation scenarios.
• To enhance feature extraction and leverage unlabeled target domain data for more robust decoding.

Main Methods:

• Developed a model-agnostic dual-attention meta-learning framework (DA-META).
• Incorporated an enhanced temporal attention module for effective feature extraction.
• Utilized a cosine similarity-based attention module to guide meta-training with unlabeled target data.
• Implemented a three-stage process: meta-task construction, guided meta-training, and fine-tuning-free meta-testing.

Main Results:

• Achieved high classification accuracies on self-collected and public datasets (e.g., 80.93% on BCI Competition IV 2b).
• Outperformed state-of-the-art methods in MI-EEG decoding.
• Demonstrated accuracy improvements across different backbone networks (EEGNet, DeepConvNet, EEG Conformer).
• Showcased superior performance in handling inter-subject variability.

Conclusions:

• The DA-META framework effectively improves MI-EEG decoding accuracy and generalization.
• The proposed attention mechanisms enhance feature extraction and target domain adaptation.
• DA-META holds significant potential for practical applications in paralysis rehabilitation and brain-computer interfaces.