Evaluation of Automated Hypnogram Analysis on Multi-Scored Polysomnographies

Area of Science:

• Sleep Medicine
• Computational Neuroscience
• Biomedical Engineering

Background:

• Automated sleep stage scoring is crucial for analyzing polysomnography data.
• Existing methods face challenges in replicating human expert variability and interpretation.
• Evaluating automated scoring requires accounting for inter-rater disagreement and prediction confidence.

Purpose of the Study:

• To develop and evaluate a novel automated sleep stage scoring method for polysomnography.
• To establish a new standard for evaluating automated sleep scoring algorithms by incorporating inter-rater variability.
• To analyze transition periods between sleep stages, a previously unstudied aspect.

Main Methods:

• A random forest model was employed to capture feature interactions and temporal dynamics.
• The model utilized features aligned with American Academy of Sleep Medicine guidelines for expert interpretability.
• A multi-labeled dataset was used, considering both inter-rater variability and prediction uncertainties for evaluation.

Main Results:

• The automated method achieved an overall accuracy of 82.7% across all epochs.
• When focusing on epochs with high certainty (63.3%), the model's accuracy reached 97.8%.
• Transition periods between sleep stages were identified and analyzed for the first time.

Conclusions:

• The proposed random forest method offers a highly accurate and time-efficient solution for automated sleep stage scoring.
• The evaluation framework, accounting for inter-rater variability, sets a new benchmark for assessing automated sleep scoring algorithms.
• The method provides scoring guidelines for medical experts, enabling manual refinement for guaranteed accuracy.