Label-efficient sleep staging from multi-channel EEG with self-supervised contrastive learning and iterative self-distillation

Abstract

Manual sleep stage classification from polysomnography (PSG) is labor-intensive and subject to expert variability, motivating automated and deployment-oriented solutions for clinical use. We present a multi-channel self-supervised learning (SSL) contrastive framework combined with iterative self-distillation for accurate and label-efficient sleep staging. The approach employs a dual-branch convolutional network that processes electroencephalogram (EEG) channels independently and integrates complementary information via a cross-attention fusion module. During pre-training, a contrastive objective leverages temporal adjacency to form positive pairs and maintains hard negatives dynamically to learn robust representations from unlabeled data. Subsequent fine-tuning with minimal labels is enhanced by iterative self-distillation through pseudo-label refinement. On the Sleep-EDF Expanded (SleepEDF-v2) dataset, the method achieves strong performance with only 1% labeled data (accuracy 76.31%, macro-F1 66.53%), competitive against existing SSL baselines. The resulting compact model and single-site training setup align with practical constraints in hospitals and wearable scenarios, reducing annotation burden and supporting secure, scalable clinical deployment.