A Two-Level Ensemble Machine Learning Framework for OSA Classification Whilst Awake from Noisy Tracheal Breathing...

Area of Science:

• Biomedical Engineering
• Respiratory Medicine
• Machine Learning

Background:

• Obstructive sleep apnea (OSA) is underdiagnosed due to complex sleep studies.
• Tracheal breathing sounds (TBS) analysis during wakefulness presents a potential screening alternative.
• Noisy environments and low-cost equipment pose challenges for acoustic OSA detection.

Purpose of the Study:

• To evaluate the feasibility of detecting OSA using TBS recorded with an inexpensive microphone in noisy conditions.
• To develop and assess a machine learning algorithm for OSA screening based on awake tracheal sounds.

Main Methods:

• Analysis of TBS from 247 individuals with varying OSA severity.
• Segmentation of breathing sounds, acoustic feature extraction, and dimensionality reduction.
• Implementation of a two-level ensemble classifier with bagged decision trees and probability voting.

Main Results:

• The algorithm achieved 77.1% accuracy, 84.3% sensitivity, and 59.9% specificity.
• Performance was lower than studies using high-quality equipment in quiet environments.
• Microphone quality significantly impacted classification performance.

Conclusions:

• Acoustic screening for OSA during wakefulness is viable even in challenging noisy environments.
• Inexpensive microphones can be used, but performance is sensitive to equipment quality.
• Further research is needed to optimize algorithms for real-world, noisy conditions.