XSleepNet: Multi-View Sequential Model for Automatic Sleep Staging.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Automating sleep staging is crucial for diagnosing sleep disorders and enabling remote monitoring.
• Current methods often rely on single data views (raw signals or time-frequency images).
• Multi-view learning for sleep staging remains challenging and underexplored.

Purpose of the Study:

• To propose XSleepNet, a novel sequence-to-sequence model for sleep staging.
• To enable joint representation learning from both raw polysomnography signals and time-frequency images.
• To adaptively manage the learning pace for each input view based on generalization and overfitting.

Main Methods:

• Developed XSleepNet, a sequence-to-sequence architecture for multi-view sleep staging.
• Implemented adaptive learning rate adjustment based on on-the-fly generalization/overfitting measures.
• Computed view-specific weights to blend gradients from different input modalities.

Main Results:

• XSleepNet effectively learns joint representations superior to single-view methods.
• The adaptive training strategy enhances robustness to varying data amounts.
• Achieved superior performance compared to single-view baselines and a simple fusion multi-view baseline.

Conclusions:

• XSleepNet outperforms existing sleep staging methods, setting a new state-of-the-art on multiple databases.
• The proposed adaptive multi-view learning approach improves sleep staging accuracy and reliability.
• This method offers a scalable solution for sleep assessment and longitudinal monitoring.