Toward Foundational Model for Sleep Analysis Using a Multimodal Hybrid-Self-Supervised Learning Framework.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Sleep Medicine

Background:

• Accurate sleep analysis is vital for health and diagnosing sleep disorders.
• Manual diagnosis of sleep data is time-consuming and subjective.
• Current automated methods often require extensive labeled datasets.

Purpose of the Study:

• To introduce SynthSleepNet, a multimodal hybrid-self-supervised learning framework for polysomnography (PSG) data analysis.
• To leverage complementary features across EEG, EOG, EMG, and ECG signals for robust sleep analysis.
• To develop an efficient temporal context module for capturing signal interdependencies.

Main Methods:

• Developed SynthSleepNet, a multimodal hybrid-self-supervised learning (SSL) framework.
• Integrated masked prediction and contrastive learning for feature representation.
• Incorporated a Mamba-based temporal context module for signal analysis.

Main Results:

• Achieved high accuracies in sleep-stage classification (89.89%), apnea detection (99.75%), and hypopnea detection (89.60%).
• Demonstrated robust performance in a semi-supervised setting with limited labels.
• Outperformed state-of-the-art methods across all evaluated downstream tasks.

Conclusions:

• SynthSleepNet offers a powerful tool for comprehensive PSG data analysis.
• The framework shows significant potential for advancing sleep disorder monitoring and diagnostics.
• The model sets a new standard for automated sleep analysis systems.