SleepTransformer: Automatic Sleep Staging With Interpretability and Uncertainty Quantification.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Deep learning models for automatic sleep scoring face challenges in clinical adoption due to "black-box" skepticism.
• Lack of interpretability hinders the trust and validation of AI-driven sleep analysis in healthcare settings.

Purpose of the Study:

• To develop an interpretable deep learning model for automatic sleep staging.
• To introduce a method for quantifying the uncertainty of AI-driven sleep scoring decisions.

Main Methods:

• Proposed SleepTransformer, a sequence-to-sequence model utilizing a transformer backbone for sleep staging.
• Developed an entropy-based method to quantify decision uncertainty, enabling deferral of low-confidence epochs to human experts.
• Visualized self-attention scores as heatmaps (epoch-level) and influence maps (sequence-level) for interpretability.

Main Results:

• SleepTransformer provides interpretable insights into sleep stage classification at both epoch and sequence levels.
• Attention scores highlight sleep-relevant EEG features and contextual epoch influences, mimicking human scoring.
• The model demonstrates performance comparable to existing methods across diverse datasets.

Conclusions:

• SleepTransformer offers interpretability and uncertainty quantification, addressing key barriers to clinical integration.
• The model shows significant potential for reliable application in clinical sleep scoring environments.