MultiConsSleepNet: Self-Supervised Contrastive Learning for a Multimodal Consistency-Based Automatic Sleep Staging...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate sleep staging is crucial for diagnosing sleep disorders and understanding sleep's role in overall health.
• Current deep learning methods for automatic sleep staging face challenges in multimodal representation extraction, leveraging signal similarities/differences, and utilizing unlabeled data.

Purpose of the Study:

• To develop a multimodal deep learning network, MultiConsSleepNet, for efficient and accurate automatic sleep staging.
• To address challenges in multimodal feature extraction, consistency learning, and leveraging unlabeled data for enhanced model generalization and practicality.

Main Methods:

• MultiConsSleepNet integrates unimodal feature extractors and a multimodal consistency extractor for electroencephalogram (EEG) and electrooculogram (EOG) signals.
• Self-supervised contrastive learning strategies are employed for both unimodal and multimodal consistency learning.
• The model is evaluated on three datasets to assess its performance with limited labeled data.

Main Results:

• MultiConsSleepNet achieves state-of-the-art performance in automatic sleep staging, particularly with limited labeled data.
• The model demonstrates high effectiveness in leveraging abundant unlabeled data to enhance practical application value.
• Achieved an average F1 score of 74.2% and accuracy of 79.6% in cross-dataset scenarios using only 10% labeled data.

Conclusions:

• MultiConsSleepNet offers a promising multimodal deep learning approach for automated sleep staging, significantly reducing the need for extensive labeled data.
• The method enhances feature consistency within and across modalities, improving model generalization and transferability.
• This approach provides a valuable tool for clinicians, improving the efficiency and reliability of sleep disorder diagnosis and clinical workflow.