Enhancing EEG Domain Generalization via Weighted Contrastive Learning

Recently, there has been notable progress in deep learning-based electroencephalogram (EEG) analysis, particularly in sleep staging classification. However, the substantial variation in EEG signals across subjects poses a significant challenge, limiting model generalization. To tackle this issue, contrastive learning-based domain generalization (DG) has been proposed and has shown promising performance. In essence, DG aims to closely associate the features of the same class across multiple domains. Throughout this process, negative pairs from different domains are pushed further away from the anchor compared to negative pairs from the same domain, leading to the emergence of domain gaps. In this paper, we propose a novel framework to balance the effects of negative samples from different domains with negative samples in the same domain. It prevents the enlargement of domain gaps and enables the extraction of subject-invariant features. For the validity of our proposed method, we experimented on the SleepEDF-78 dataset. Experimental results demonstrated that our method outperformed the previous methods considered in our experiments.