Sleep-disordered breathing (SDB) is a serious and prevalent condition that results from the collapse of the upper airway during sleep, which leads to oxygen desaturations, unphysiological variations in intrathoracic pressure, and sleep fragmentation. Its most common form is obstructive sleep apnoea (OSA). This has a big impact on quality of life, and is associated with cardiovascular morbidity. Polysomnography, the gold standard for diagnosing SDB, is obtrusive, time-consuming and expensive. Alternative diagnostic approaches have been proposed to overcome its limitations. In particular, acoustic analysis of sleep breathing sounds offers an unobtrusive and inexpensive means to screen for SDB, since it displays symptoms with unique acoustic characteristics. These include snoring, loud gasps, chokes, and absence of breathing. This thesis investigates deep learning methods, which have revolutionised speech and audio technology, to robustly screen for SDB in typical sleep conditions using acoustics. To begin with, the desirable characteristics for an acoustic corpus of SDB, and the acoustic definition of snoring are considered to create corpora for this study. Then three approaches are developed to tackle increasingly complex scenarios. Firstly, with the aim of leveraging a large amount of unlabelled SDB data, unsupervised learning is applied to learn novel feature representations with deep neural networks for the classification of SDB events such as snoring. The incorporation of contextual information to assist the classifier in producing realistic event durations is investigated. Secondly, the temporal pattern of sleep breathing sounds is exploited using convolutional neural networks to screen participants sleeping by themselves for OSA. The integration of acoustic features with physiological data for screening is examined. Thirdly, for the purpose of achieving robustness to bed partner breathing sounds, recurrent neural networks are used to screen a subject and their bed partner for SDB in the same session. Experiments conducted on the constructed corpora show that the developed systems accurately classify SDB events, screen for OSA with high sensitivity and specificity, and screen a subject and their bed partner for SDB with encouraging performance. In conclusion, this thesis makes promising progress in improving access to SDB diagnosis through low-cost and non-invasive methods.
