AI-based classification of OSA severity in awake subjects using breathing sound analysis

Abstract

Obstructive sleep apnea (OSA) is a prevalent, underdiagnosed disorder associated with cardiovascular and cognitive complications. The diagnostic gold standard, polysomnography (PSG), is costly and resource-intensive, motivating the development of accessible screening alternatives. This thesis proposes that tracheal breathing sounds (TBS) recorded during wakefulness, combined with anthropometric data, can predict OSA severity using three AI-based approaches: classification, feature interpretation, and non-invasive clinical screening. The first approach involves developing a six-pairwise base-model framework using TBS combined with anthropometric data. A multidomain feature set comprising spectral, temporal, time-frequency, and nonlinear features was extracted. Features were selected using statistical filtering, SHAP ranking, and Recursive Feature Elimination before training pairwise binary classifiers. Nonlinear and spectral features were the most discriminative indicators of the severity of upper airway obstruction. The second approach extends the base-models framework by introducing a stacked meta-model architecture for multi-class OSA severity prediction across both three-class and four-class stratification schemes. A conformal prediction (CP) framework was further incorporated to construct adaptive prediction sets with guaranteed statistical coverage, thereby transparently communicating diagnostic uncertainty in ambiguous cases. Stacking pairwise classifiers achieved clinically meaningful multi-class severity stratification. Conformal prediction provided principled uncertainty quantification for point-of-care triage. The third approach extends the base-models framework by focusing on the clinical interpretability and physiological relevance of the extracted acoustic features. Features were evaluated using discrimination, stability, anthropometric correlation, and SHAP explainability. A Structure–Function–Symptom framework linked discriminative acoustic descriptors to airway anatomy, airflow turbulence, and AHI-defined disease burden. The results demonstrated that spectral, bispectral, and fractal-based nonlinear features consistently emerged as the most stable and physiologically meaningful biomarkers, linking TBS signatures of wakefulness to progressive upper airway dysfunction. Together, these approaches form a wakefulness-based framework for OSA screening and severity assessment using TBS recorded in under 10 minutes. For binary OSA screening, the deep learning model achieved 74.9% accuracy, 76.1% sensitivity, and 73.3% specificity. For pairwise severity discrimination using the base models, the best binary classifier (Non-OSA vs. Severe-OSA) reached 88.2% accuracy, 83.3% sensitivity, and 90.9% specificity. The meta-modeling framework further achieved 76.7% accuracy, 77.7% sensitivity, and 87.1% specificity in the three-class setting, and 76.7% accuracy, 75% sensitivity, and 92% specificity in the four-class setting. Conformal prediction further provides calibrated confidence sets, enhancing clinical trustworthiness and uncertainty-aware point-of-care triage.