Sound-based sleep assessment with controllable subject-dependent embedding using Variational Domain Adversarial Neural Network

Abstract

Sleep quality assessment as an indicator of daily health care plays an important role in our physiological and mental activity. Sound during sleep contains rich information on biological activities, such as body movement, snoring, and sleep bruxism. However, sound features differ depending on individual and environmental differences. In order to develop a wide-rage applicable daily sleep assessment, this paper utilizes deep learning to ease individual and environmental differences of sound features. Firstly, by Variational Domain Adversarial Neural Network (VDANN) encodes sound events into latent representation, simultaneously eliminates subject-dependent features. Then, sleep pattern in the obtained latent space is trained by Long Short-Term Memory (LSTM) with associated sleep assessment of one night. We performed age group estimation from normal sleep as an objective indicator of sleep comparing to their age group. The experiment with more than 100 subjects showed that VDANN is able to extract subject independent features, and the proposed method outperforms the conventional method for age group estimation from sleep sound even for new subjects. In addition, our model is able to personalize by controlling subject-dependent embedding when after data accumulation of the subject.