CSANet: enhancing cross-subject EEG classification with contrastive subject adaptation

Electroencephalogram (EEG)-based cross-subject classification remains challenging due to inter-subject variability in neural signal patterns. While deep learning approaches have shown promise, their performance degrades significantly when applied to unseen subjects. This paper proposes a novel Contrastive Subject Adaptation network (CSANet) to enhance cross-subject generalization by learning subject-invariant yet task-discriminative EEG representations. Unlike conventional contrastive learning that relies on self-supervised instance discrimination, CSANet incorporates explicit class labels to construct positive and negative pairs, enforcing intra-class compactness and inter-class separation across different subjects. We design a hybrid loss function combining supervised contrastive loss, effectively minimizing both feature distribution divergence and classification error. Comprehensive experiments on three public datasets demonstrate that CSANet achieves state-of-the-art performance, improving mean cross-subject classification accuracy compared to baseline methods. Ablation studies validate the synergistic effects of supervised contrastive learning and domain adaptation components.