Feature alignment and enhancement network with guided tuning for non-stationary EEG classification

Objective.Electroencephalogram (EEG) signal variability caused by external factors and subject differences limits the adaptation of motor imagery (MI) classification models in brain-computer interfaces (BCIs). Existing domain alignment methods often inadequately utilize critical source and target domains information, leading to negative transfer problems. This paper proposes a Feature Alignment and Enhancement Framework for cross-domain MI-EEG classification to address these limitations.Approach.First, by aligning the covariance matrices of the source and target domains, the spatial distributions of the two domains are preliminarily aligned, establishing a consistent foundation for feature mapping. Second, a conditional domain adversarial network optimizes cross-domain representations, reducing distribution discrepancies while enhancing discriminability. Finally, this paper introduces an EEG feature-based guided tuning method. This method extracts high-confidence features from both the source and target domains and generates centroid features to construct cross-domain feature banks. The input feature representations are dynamically optimized by attending to the relationships between centroid features, thus enhancing the model's adapt-ability to target domain tasks.Main results.Experimental data show that in the four-class MI task of the BCI Competition IV-2a dataset, the cross-session and cross-subject model classification accuracies were 76.89% and 57.91%, respectively. The model achieved accuracy rates of 84.61% and 82.78% on the BCI Competition IV-2b datasets and the High Gamma Datasets, respectively, as well as 84.09% and 70.81%.Significance.The proposed framework effectively mitigates cross-domain variations, providing a reliable solution for cross-session and cross-subject MI-EEG classification.