Attention on Sleep Stage Specific Characteristics.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Manual sleep stage classification is time-consuming and relies on visual inspection of polysomnography data.
• Deep neural networks offer a faster alternative but often lack transparency, making it difficult to understand their decision-making process.
• Ambiguity in sleep staging arises from coarse labeling of 30-second windows, potentially mixing class-specific features.

Purpose of the Study:

• To enhance the transparency of deep neural classifiers for sleep staging.
• To develop a dynamic discrete attention module for identifying relevant input data subsets.
• To utilize the attention mechanism as a tool for discovering sleep stage-specific features in polysomnography data.

Main Methods:

• Proposed a dynamic discrete attention module compatible with standard classification networks.
• Validated the method on both synthetic and patient polysomnography data.
• Analyzed dynamic attention masks to assess channel selection adaptivity.

Main Results:

• Demonstrated that a small subset of data within a 30-second window is sufficient for accurate sleep classification.
• Observed performance enhancement due to the proposed attention mechanism.
• Showcased clear, sleep stage-adaptive channel selection through analysis of attention masks.

Conclusions:

• The dynamic discrete attention module significantly improves transparency in deep neural network-based sleep staging.
• The method effectively identifies crucial data segments and aids in discovering novel sleep stage-specific features.
• Attention mechanisms offer a powerful approach for interpretable and accurate sleep analysis using polysomnography.