Self-supervised Learning for Encoding Between-Subject Information in Clinical EEG

Progress in deep learning for the analysis of clinical EEG data has been hindered by label noise and labeled data sample sizes. While self-supervised learning (SSL) offers a promising solution by learning representations without labels, their practical utility for clinical applications remains poorly understood. Through systematic evaluation using two large clinical EEG datasets, we provide a comprehensive assessment of SSL for pathology detection while controlling for demographic confounds. We introduce a novel yet simple contrastive learning approach that explicitly encodes between-subject information, achieving superior detection of both neurological and psychiatric pathology compared to existing methods. We use data subsampling to highlight differences in the dynamics of representation learning between the neurological and psychiatric domain. Our evaluations help characterize the strengths and limitations of current SSL approaches, thereby providing guidance for applying and developing SSL methods for clinical EEG.