Predicting Subjective Sleep Quality Using Recurrent Neural Networks

Abstract

Our goal is to predict subjective sleep quality (SSQ) from objective sleep data and identify the causes and markers of the variances within “normal” sleep. Such information would increase our understanding of the causes of variation in SSQ and potentially improve our ability to improve SSQ. Previous approaches rely on human annotation of the electroencephalographic (EEG) brain signals, to deal with the noisy, high dimensional nature of the EEGs. We aim to use recurrent neural networks to directly analyze and extract useful information from EEG brain signals. We analyze population-based overnight sleep polysomnography data obtained from 4885 community-dwelling adults. We use convolutional and recurrent neural networks to process the EEGs and combine them with information related to health and lifestyle to predict subjective depth and restfulness of sleep. We compare the coefficient of determination to the ones obtained with regression methods and technician annotations of the EEGs in previous studies. Predicting SSQ from our data set of community-dwelling adults using RNNs to analyze the whole EEG signals appear to be less accurate than previous approaches predictions. It might be necessary to acquire more data, possibly with new variables that might be better correlated with SSQ. RNNs are, however, able to extract variables correlated with SSQ from EEG signals. Our results provide insights into how RNNs can be used to extract information from brain signals and how methods such as hierarchical clustering analysis can help neural networks predict subjective variables from polysomnography data.