Noncontact Sleep Study Based on an Ensemble of Deep Neural Network and Random Forests

Sleep quality, which is an undervalued health issue that affects well-being and daily lives, is checked through the polysomnography (PSG), considered as the gold standard for determining sleep stages. Due to the obtrusiveness of its sensor attachments, recent sleep stage classification algorithms using noninvasive sensors have been developed and commercialized. However, the newly developed devices and algorithms used in the previous studies have lacked the detection of non-rapid eye movement and rapid eye movement sleep, which are known to be correlated with the development of sleep disorders, cardiovascular disease, metabolic disease, and neurodegeneration. We devise a novel approach to employ ensemble of deep neural network and random forest for the performance of noncontact sleep stage classification. Notably, this paper is designed based on the PSG data of sleep-disordered patients, which were received and certified by professionals at Hanyang University Hospital. The efficiency of the proposed algorithm is highlighted by contrasting sleep stage classification performance with previously proposed methods and a commercialized sleep monitoring device called ResMed S+. The proposed algorithm was assessed with random patients following gold-standard measurement schemes (PSG examination), and results show a promising novel approach for determining sleep stages in an economical and unobtrusive manner.