Deep representation learning with cross attention-based multi-feature fusion for sleep apnea detection using sleep respiratory sound

Obstructive sleep apnea (OSA) is a prevalent but often underdiagnosed sleep disorder linked to several health risks. While polysomnography (PSG) remains the clinical gold standard for OSA diagnosis, it is costly, time-consuming, and uncomfortable for patients. Automated OSA detection systems based on machine learning offer a promising alternative, but most existing approaches rely on invasive physiological signals (e.g., electrocardiogram) or require specialized equipment, limiting their practicality. This study aims to develop a non-invasive, cost-effective deep learning framework for OSA detection using sleep-related respiratory sound signals. We introduce a novel latent representation learning method to extract meaningful latent features from sleep sound signals using a supervised signal-guided encoder–decoder block. A hierarchical cross-attention fusion block is then proposed to integrate the learned representations with handcrafted acoustic features and demographic information. Finally, a TabNet classifier is used to perform apnea–hypopnea event classification, leveraging an adaptive feature selection and transformation mechanism. Extensive experiments on clinical sleep laboratory data demonstrate that our model significantly outperforms existing baselines in both segment-based apnea event detection and OSA severity classification. This work provides a non-invasive, cost-effective solution for OSA detection, enabling large-scale screening and facilitating early diagnosis and intervention in clinical settings. The complete implementation of our method is publicly available at: https://github.com/pw220/osa-audio.