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

## TL;DR

A machine learning framework using awake breathing sounds can detect sleep apnea with reasonable accuracy in noisy environments.

## Contribution

A two-level ensemble method using anthropometric stratification and probability-based voting improves OSA detection from noisy awake breathing sounds.

## Key findings

- The framework achieved 77.1% accuracy, 84.3% sensitivity, and 59.9% specificity in noisy conditions.
- Stratifying classifiers by anthropometric profiles improved robustness in real-world noise.
- Microphone quality significantly affects classification performance in acoustic OSA detection.

## Abstract

What are the main findings?
A two-level ensemble machine learning framework using awake tracheal breathing sounds achieved 77.1% accuracy, 84.3% sensitivity, and 59.9% specificity in a noisy recording environment.Stratifying nine sub-classifiers by anthropometric profiles and aggregating predictions using probability-based voting improved robustness under real-world noise.

A two-level ensemble machine learning framework using awake tracheal breathing sounds achieved 77.1% accuracy, 84.3% sensitivity, and 59.9% specificity in a noisy recording environment.

Stratifying nine sub-classifiers by anthropometric profiles and aggregating predictions using probability-based voting improved robustness under real-world noise.

What are the implications of the main findings?
The results support feasible awake OSA screening in noisy, real-world recordings, motivating further validation on larger and more balanced cohorts.The framework’s high sensitivity suggests its effectiveness as a primary screening tool to prioritize at-risk patients for confirmatory clinical diagnosis.

The results support feasible awake OSA screening in noisy, real-world recordings, motivating further validation on larger and more balanced cohorts.

The framework’s high sensitivity suggests its effectiveness as a primary screening tool to prioritize at-risk patients for confirmatory clinical diagnosis.

Obstructive sleep apnea (OSA), defined by repetitive airway obstruction during sleep, is significantly underdiagnosed, mainly due to the resource-intensive and time-consuming nature of sleep assessment technologies. Machine learning analysis of the tracheal breathing sounds (TBS) whilst awake offers an alternative approach for OSA quick screening. This study aimed to address the challenge of wakefulness OSA detection using TBS recorded with an inexpensive microphone in a noisy environment. Data of 247 individuals with various degrees of OSA severity were analyzed. Recorded data were segmented into inspiration and expiration phases, followed by acoustic features extraction, feature reduction, and classification. A two-level ensemble architecture was implemented. Nine sub-classifiers were stratified by anthropometric profiles. Each sub-classifier was constructed as an ensemble of bagged decision trees, with a final prediction via probability-based voting. The proposed algorithm achieved an accuracy of 77.1%, sensitivity of 84.3%, and specificity of 59.9%. Although these results have lower performance than those obtained previously using a high-quality microphone in a quiet room, they demonstrate that acoustic OSA detection whilst awake remains feasible, even in very noisy environments. Nevertheless, microphone quality emerged as a key determinant of classification performance.