Learning to Detect Sleep Micro-Events From Coarse Sleep Stage Annotations.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Sleep Medicine

Background:

• Sleep micro-events like spindles and K-complexes are vital for cognitive functions.
• Current AI detection methods need extensive expert-labeled data, which is scarce.
• Inter-rater variability in expert annotations introduces inconsistencies.

Purpose of the Study:

• To develop an automated sleep micro-event detection system using readily available coarse sleep stage annotations.
• To overcome the data bottleneck in supervised AI models for sleep analysis.
• To infer expert logic from coarse annotations while handling annotation inconsistencies.

Main Methods:

• Introduction of the Weakly Supervised Sleep Micro-Event Detector (WSSMED).
• WSSMED utilizes a dual-branch architecture with wave prototype and cluster modules.
• Leverages coarse sleep stage annotations for training, inferring fine-grained patterns.

Main Results:

• WSSMED achieved state-of-the-art performance in detecting sleep spindles and K-complexes.
• Achieved high precision, recall, and F1-scores at both sample and event levels.
• Subject-level analysis showed WSSMED's micro-event metrics closely align with expert annotations.

Conclusions:

• WSSMED effectively detects sleep micro-events using weakly supervised learning.
• The method mitigates performance issues from annotation variability.
• Demonstrates significant clinical potential for reliable sleep micro-event analysis and cognitive function assessment.