Interpretable Acoustic Features from Wakefulness Tracheal Breathing for OSA Severity Assessment

Area of Science:

• Biomedical Engineering
• Data Science
• Sleep Medicine

Background:

• Obstructive Sleep Apnea (OSA) is a common sleep disorder linked to serious health issues.
• Current diagnosis via polysomnography (PSG) is costly and inaccessible.
• There is a need for simpler, non-invasive OSA diagnostic tools.

Purpose of the Study:

• To develop and validate a machine learning framework for classifying OSA severity.
• To utilize non-invasive tracheal breathing sounds (TBS) and anthropometric data.
• To differentiate between four OSA severity levels: non, mild, moderate, and severe.

Main Methods:

• Collected TBS and anthropometric data from 199 participants.
• Preprocessed and extracted multi-domain features from TBS.
• Employed a three-stage feature selection process (univariate, SHAP, RFE).
• Trained and validated ensemble learning models using bootstrap aggregation and cross-validation.

Main Results:

• The framework accurately discriminated between OSA severity groups.
• Combining TBS and anthropometric features improved classification performance.
• Audio biomarkers from TBS demonstrated efficacy for OSA screening.

Conclusions:

• A TBS and anthropometric data-based model is a promising alternative/supplement to PSG for OSA severity detection.
• This non-invasive approach enhances screening scalability and accessibility.
• The method may facilitate earlier detection of undiagnosed OSA cases.