Accessible Obstructive Sleep Apnea Screening Using Classical Acoustic Speech Representations

Obstructive sleep apnea (OSA) is a chronic respiratory disorder characterized by recurrent interruptions in breathing during sleep. The gold standard for clinical OSA diagnosis is the polysomnography test, which is a rather cumbersome and expensive procedure. As a result, other alternatives for screening OSA have gained attention. Speech, an accessible modality, shares similar anatomical structures that contribute to OSA. This study proposes a novel pipeline based on classical acoustic features to estimate the risk of OSA using five vowels and two phonemes recorded in standing and sitting postures. A transfer learning approach that incorporates magnitude and phase-based representations alongside a pre-trained SEResNet-50 was used to compare with our model. Our proposed framework achieved a remarkable F1-score of 0.93 in classifying 35 subjects living in homeless shelters, proving the feasibility of employing acoustic models in accessible speech-based OSA screening.