Automatic Segmentation to Cluster Patterns of Breathing in Sleep Apnea

Area of Science:

• Sleep Medicine
• Data Science
• Biomedical Engineering

Background:

• Polysomnography (PSG) recording annotation for obstructive sleep apnea (OSA) diagnosis is crucial but resource-intensive.
• Manual analysis of PSG data is time-consuming and requires specialized clinical expertise.
• Developing automated methods can significantly improve the efficiency and accessibility of OSA diagnosis.

Purpose of the Study:

• To develop a data-driven, unsupervised hierarchical clustering algorithm for detecting and visualizing respiratory events in PSG recordings.
• To create a model for OSA-related event detection that is independent of manual annotations.
• To evaluate the algorithm's performance in identifying breathing patterns indicative of OSA.

Main Methods:

• An unsupervised hierarchical clustering approach was applied to PSG recordings.
• The algorithm focused on detecting and visually presenting breathing patterns.
• Evaluation utilized 10 recordings from the Sleep Heart Health Study, comparing algorithm output to manual annotations.

Main Results:

• The algorithm achieved an F1-score of 0.58 in detecting respiratory events versus no events.
• It demonstrated 100% accuracy in predicting the presence of OSA based on an apnea-hypopnea index (AHI) ≥ 15.
• Potential reasons for the F1-score variation include event placement precision and scoring variability.

Conclusions:

• The unsupervised clustering method serves as a proof of concept for detecting and visualizing OSA-related breathing patterns.
• Despite challenges, the algorithm showed strong diagnostic agreement for OSA.
• Further improvements are possible to enhance the F1-score for event detection while maintaining diagnostic accuracy.