Prototype-based sleep micro-structure learning for explainable and robust multimodal recognition of sleep-related...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Sleep disturbances are linked to reduced quality of life and neurodegenerative diseases.
• Automated sleep staging networks lack transparency, hindering clinical trust and biomarker discovery.

Purpose of the Study:

• Introduce ProtoSleepNet (PSN), a novel prototype-based sequence-to-sequence architecture for interpretable sleep staging.
• Validate PSN's accuracy, robustness, and ability to identify disease-specific sleep alterations.

Main Methods:

• Developed PSN, a deep learning model using learned prototypes to represent sleep microstructure.
• Validated PSN on over 10,000 recordings across 10 datasets, testing robustness to channel occlusion.
• Developed an explainability framework to translate prototypes into clinically relevant rules.

Main Results:

• PSN achieved human-level sleep staging accuracy, comparable or superior to state-of-the-art methods.
• Demonstrated robustness against channel occlusion attacks.
• Prototype sequences (prototype-grams) differentiated Parkinson's and Alzheimer's patients from healthy controls without disease-specific training.

Conclusions:

• PSN provides an interpretable alternative to black-box sleep staging models.
• Learned prototypes capture clinically meaningful sleep microstructure patterns.
• Prototype-grams can serve as biomarkers for neurodegenerative diseases, aligning with known pathophysiology.