Developing REM Sleep Prediction Models Using Smart Home Sensor Data

In this study, we discuss the development of a model for predicting the duration of REM (rapid eye movement) sleep utilizing sensor data from smart homes. It has been noted that a shortage of REM sleep is associated with increased cardiovascular strain and a heightened risk of overwork-related fatalities. Therefore, it is crucial to provide users with a system that offers advice on their daily sleep, activity, and environmental habits to obtain high-quality sleep. However, collecting sleep data continuously for an extended period is impractical. We aim to develop a reliable model that predicts the duration of REM sleep using sensor data with missing values collected from smart homes. The model utilizes data on the users' sleep, activity, and environmental conditions for several days. This study uses Boruta as the feature selection method, adopts Random Forest as the machine learning technique, and utilizes LOPO-CV and MAE as evaluation methods. The predictive accuracy of the proposed model surpassed that of models trained solely on historical REM sleep data or those without any variable selection. Further examining the relationship between selected features and REM sleep may lead to developing a personalized system that can enhance the quality of users' sleep.