MtRBD: Advancing iRBD Analysis With Multi-Task Learning for Joint Sleep Staging and RSWA Detection.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Technology

Background:

• Rapid eye movement (REM) sleep without atonia (RSWA) is crucial for diagnosing REM sleep behavior disorder (RBD).
• Current manual annotation methods are time-consuming and prone to variability.
• Existing automated methods struggle with data imbalance and multichannel signal fusion, and do not fully leverage the relationship between sleep staging and RSWA detection.

Purpose of the Study:

• To develop an advanced automated framework for improved sleep staging and RSWA detection.
• To address limitations of current automated methods, including data imbalance and complex signal fusion.
• To enhance the clinical interpretability and reliability of automated RBD detection.

Main Methods:

• Proposed a multi-task learning framework, the Multi-task RBD Network (MtRBD), integrating a Multi-scale Information Attention Bottleneck Network (MIABNet) backbone.
• MIABNet enhances multichannel signal fusion, while MtRBD employs dynamic feature enhancement for cross-task information flow.
• Utilized modality selection and Grad-CAM visualization for enhanced clinical interpretability.

Main Results:

• Achieved 83.0% accuracy for sleep staging and 93.6% for RSWA detection on the CZ-RBD dataset.
• Demonstrated 77.6% accuracy for critical RSWA event identification.
• Outperformed single-task methods in cross-subject validation, confirming robustness.

Conclusions:

• The proposed MtRBD framework offers a significant advancement in automated RSWA detection and sleep staging.
• This technology provides reliable support for the early detection of RBD and associated neurodegenerative diseases.
• The multi-task approach effectively addresses data imbalance and enhances multichannel signal fusion for improved diagnostic accuracy.