Direct Quantification of Uncertainty in Deep Learning-Based Automatic Sleep Staging.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Deep learning models automate sleep staging but lack transparency regarding uncertainty.
• Quantifying uncertainty is crucial for clinical adoption and reliable decision-making.

Purpose of the Study:

• To compare methods for quantifying uncertainty in deep learning-based automatic sleep staging.
• To evaluate the effectiveness of a novel Hypnodensity Interval (HI) method.

Main Methods:

• Three models were analyzed: traditional thresholding, Monte Carlo (MC) dropout with mean/standard deviation thresholds, and the novel HI method.
• Models were trained on the STAGES dataset and evaluated on the DOD dataset by removing uncertain epochs.

Main Results:

• All models achieved >83% accuracy; MC dropout with mean thresholding showed the best performance improvement (92% accuracy).
• The HI method effectively identified uncertainty, particularly in N2/N3 stage misclassifications, performing comparably to traditional methods.

Conclusions:

• Uncertainty in automatic sleep staging can be reliably quantified from hypnodensity outputs.
• The HI method provides a flexible and justifiable approach for uncertainty assessment.
• Findings support clinical integration by enhancing transparency and enabling targeted manual review.