Single-source unsupervised domain adaptation for cross-subject MI-EEG classification based on discriminative information

Electroencephalography (EEG) provides a wealth of physiological and psychological information. Decoding EEG signals enables machines to recognize brain activity, a crucial aspect in brain-computer interaction and medical rehabilitation. However, the non-stationarity and inter-individual variability of EEG signals pose challenges for existing EEG signal classification models to achieve the desired level of cross-subject generalization, limiting the practical applications of EEG-based brain computer interface systems. Unsupervised Domain Adaptation (UDA) aims to improve the model’s generalization performance on the target domain by minimizing the discrepancies between the source and target domain data distributions. Many researchers treat different subjects as distinct domains and utilize Unsupervised Domain Adaptation (UDA) in transfer learning to guide the model for effective cross-subject EEG Classification. Strategies involve reducing distribution discrepancies between the source and target domains by minimizing well-designed discrepancy metrics or using an adversarial discriminator to capture domain-invariant features. However, neglecting category-discriminatory (abbreviated as discriminative) features leads to the limited effectiveness of these domain-level alignment methods. To tackle these issues, we propose the Discriminative Clustering Domain Adaptation Network (DCDAN), aimed at utilizing an unsupervised approach to construct discriminative information for facilitating Unsupervised Domain Adaptation (UDA) in achieving category-level feature alignment between the source and target domains. Specifically, we employ a clustering algorithm based on adversarial domain adaptation to associate pseudo-labels with target domain samples. Implementing a self-supervised training framework enables the model to acquire discriminative features. Moreover, we introduce the Pseudo Label-Common Spatial Pattern (PL-CSP) component, which integrates the prediction confidence of pseudo-labels from target domain samples with discriminative information from source domain samples through a covariance matrix weighting strategy based on sample confidence. This integration enhances the robustness of the learned spatial filter on the target domain. Experimental results on public datasets demonstrate the effectiveness of our proposed method in extracting more discriminative features.