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

Motor imagery electroencephalography (MI-EEG) decoding demonstrates significant potential for paralysis rehabilitation, and its generalization capability is often compromised by intersubject variability and scarcity of labeled target domain data. Meta-learning has emerged as a promising approach for unsupervised domain adaptation problem. However, existing implementations suffer from two critical limitations: insufficient feature extraction and overlooking the guiding role of unlabeled target data. To overcome these challenges, we propose a dual-attention meta-learning framework (DA-META) with model-agnostic architecture in this paper. The framework comprises three stages: meta-task construction, guided meta-training, and fine-tuning-free meta-testing. In the guided meta-training stage, DA-META incorporates two key attention mechanisms: an enhanced temporal attention module for effective feature extraction, and a cosine similarity-based attention module to leverage the guidance of target domain. Using EEGNet as the backbone network, DA-META achieves mean classification accuracies of 68.04% and 76.61% on self-collected datasets from patients and healthy subjects, and 73.29% and 80.93% on the public BCI Competition IV 2a and 2b datasets, outperforming state-of-the-art methods. When employing EEGNet, DeepConvNet, and EEG Conformer as backbone networks respectively, the framework achieves accuracy improvements of 5.17%, 2.56%, and 0.85% on the 2a dataset, compared to the baseline. These results demonstrate the framework's superior ability to handle inter-subject variability and its significant potential to improve practical applicability.