Search

Analysis of sleep stages using electroencephalograms (EEGs), a critical procedure in health monitoring, has been researched extensively. Scoring of the sleep stages is highly dependent on experts’ knowledge or hand-crafted features created by experts. Recent advancements in deep learning have, nevertheless, successfully automated the scoring process with accuracy comparable to human performance. Previous methods were less convenient as they employed specific EEG devices designed for the experiments. Such devices are usually expensive and require a special environment and arrangements for their use. In this study, we propose an automatic sleep-stage scoring system using a patch-type wearable EEG sensor based on deep learning. The proposed sensor, namely Haru, is lightweight, inexpensive, and easy to use. As for the training architecture, we implement a model based on the state-of-the-art deep learning algorithm specifically designed for automatic sleep staging. We record three channels of sleep EEGs from 30 subjects at a sampling rate of 250 Hz with the proposed system and a polysomnogram (PSG) device simultaneously. Training datasets are created on the basis of the annotated samples in PSG measurement. In the experiment, we train our model and evaluate the results using the leave-one-out cross-validation method. The results indicate that the proposed Haru Sleep is capable of scoring sleep stages with a scoring accuracy of 78.6% and an F1 score of 73.4%, which is comparable to that obtained by clinical PSG devices.