Quasi-supervised scoring of human sleep in polysomnograms using augmented input variables.

Area of Science:

• Neuroscience
• Computational Biology
• Biomedical Engineering

Background:

• Manual sleep scoring is labor-intensive and prone to errors due to human variability.
• Existing automated sleep scoring algorithms are either supervised (requiring extensive labeled data) or unsupervised (less accurate).

Purpose of the Study:

• To develop a quasi-supervised sleep scoring classifier that leverages available human scores while accommodating data limitations.
• To improve the accuracy and efficiency of automated sleep stage classification.

Main Methods:

• Extracted EEG, EMG, and EOG features from polysomnograms in 30s epochs.
• Developed a quasi-supervised framework using Gaussian mixtures and hidden Markov models, augmenting unsupervised models with partial human scoring.
• Simulated various scoring scenarios, including incomplete state scoring and inter-rater variability.

Main Results:

• The quasi-supervised classifier significantly outperformed unsupervised models in accuracy.
• Performance was comparable to fully supervised models, despite using only partial scoring data.
• The algorithm effectively mimicked human scoring patterns while mitigating human limitations.

Conclusions:

• Quasi-supervised learning offers a promising approach for automated sleep scoring.
• This method enhances accuracy and efficiency by intelligently utilizing limited human expertise.
• The developed algorithm provides a robust solution for objective sleep analysis.