Iterative expert-in-the-loop classification of sleep PSG recordings using a hierarchical clustering

Area of Science:

• Sleep Medicine
• Biomedical Engineering
• Signal Processing

Background:

• Manual sleep scoring of polysomnography (PSG) is subjective and time-consuming.
• Existing automatic classifiers lack clinical confidence, leading to continued reliance on manual methods.
• There is a need for objective and efficient PSG data evaluation.

Purpose of the Study:

• To develop a semi-supervised, data-driven approach for objective and efficient PSG data evaluation.
• To improve the accuracy and speed of sleep signal classification.

Main Methods:

• A semi-supervised algorithm identifies representative signal segments for manual scoring by a neurologist.
• The algorithm then automatically classifies the remaining recording segments using these scored templates.
• The method is evaluated on 36 PSG recordings from insomniacs and healthy controls.

Main Results:

• The developed method provides faster and more objective PSG data evaluation compared to manual scoring.
• Classification accuracy is improved by approximately 14% over existing automated classifiers.
• Comparable classification results were achieved across both patient and control datasets.

Conclusions:

• The proposed semi-supervised method offers a novel approach to sleep PSG data processing.
• Its performance is comparable to existing automated methods, with accuracies ranging from 75-95%.
• The technique is versatile, applicable across different recording devices, standards, and physiological signals.