A Multimodal Consistency-Based Self-Supervised Contrastive Learning Framework for Automated Sleep Staging in Patients...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automated sleep staging is crucial but faces challenges like limited data and poor generalization.
• Existing deep learning methods struggle with multimodal feature extraction and application to specific patient groups like those with disorders of consciousness (DOC).

Purpose of the Study:

• To develop a multimodal consistency-based sleep staging network (MultiConsSleepNet) that addresses limitations in current deep learning approaches.
• To improve sleep staging accuracy and generalizability, particularly for patients with DOC, by effectively utilizing limited labeled and abundant unlabeled data.

Main Methods:

• Proposed MultiConsSleepNet utilizes unimodal and multimodal feature extractors for electroencephalograms (EEGs) and electrooculograms (EOGs).
• Incorporated self-supervised contrastive learning strategies for both unimodal and multimodal consistency learning to leverage unlabeled data.
• Focused on exploring universal representations and intra- and inter-modal feature consistency.

Main Results:

• MultiConsSleepNet achieved state-of-the-art performance on public sleep staging datasets with limited labeled data.
• The model demonstrated effective utilization of unlabeled data, enhancing practical applicability.
• Promising results were observed on a self-collected dataset of patients with DOC, indicating potential for clinical application.

Conclusions:

• MultiConsSleepNet offers an effective solution for sleep staging, especially when labeled data is scarce.
• The model's ability to generalize and its promising performance on DOC patients provide a new avenue for sleep research in this population.