0950 Sound-based Sleep Staging at Home Using Smartphone via Deep Learning

Abstract Introduction Daily sleep tracking at home is growing in demand as more and people are aware of the significance sleep. The objective this study to propose a sound-based staging model based on deep learning that works well environments with recorded audio data from general smartphones. Methods Three different datasets were used. A labeled hospital dataset (PSG audio, N=812) an unlabeled (audio only, N=829) used for training. limited number sound N=45) testing. Our proposed HomeSleepNet has three components: (1) supervised using trains make correct predictions environments; (2) unsupervised domain adaptation (UDA), which both data, transferred power by adversarial training; (3) augmentation consistency training (UDC), augmented adding noise trained consistent original data. After all training, expected robust despite presence. Results achieved 76.2% accuracy task 3-stage classification case (Wake, NREM, REM). Specifically, it correctly predicted 63.4% wake, 83.6% NREM sleep, 64.9% REM contributions UDA UDC demonstrated following results. without was 69.2%. Either addition or improved performance, increased 69.3% 73.5% UDC. As expected, (i.e., HomeSleepNet) best 7% increase compared components. Conclusion To our knowledge, first conducted environments. Moreover, sounds commercial smartphones not through professional devices. introduced reliable convenient method daily home. Support (if any)