Efficient Sleep Staging With Bayesian Uncertainty-Guided Active Learning.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Automated sleep staging is crucial for large-scale and home-based sleep monitoring.
• Manual sleep annotation by experts is time-consuming, labor-intensive, and limits clinical adoption of automated systems.
• Existing automated systems lack reliable adaptability to new subjects, necessitating expert review.

Purpose of the Study:

• To develop an adaptive and efficient sleep staging system that reduces annotation workload and preserves expert-level accuracy.
• To introduce BayesSleepNet, a novel framework integrating Bayesian uncertainty quantification with active learning for adaptive sleep staging.
• To quantify model uncertainty using principled Bayesian modeling and Monte Carlo sampling.

Main Methods:

• BayesSleepNet employs Bayesian modeling with distributions over network weights and Monte Carlo sampling for uncertainty quantification.
• A two-stage sample selection strategy utilizes uncertainty estimates to fine-tune the model with representative epochs and prioritize uncertain samples for expert review.
• The framework was evaluated across four public sleep datasets.

Main Results:

• BayesSleepNet achieved performance improvements of 7.60% in accuracy, 8.27% in macro-F1, and 0.104 in Cohen's kappa.
• The system required manual annotation of only 20% of data from new subjects.
• BayesSleepNet is computationally lightweight with fewer parameters than state-of-the-art models.

Conclusions:

• Uncertainty-aware active learning offers a practical and cost-efficient paradigm for semi-automated sleep staging.
• BayesSleepNet demonstrates significant clinical promise for reducing annotation workload while maintaining high accuracy.
• The adaptive nature and computational efficiency of BayesSleepNet support its potential for widespread clinical adoption.