A Smart IoT System for Continuous Sleep State Monitoring

Sleep is a vital process in the maintenance of health and well-being. The effects of poor or deprived sleep can have a major impact on the daily, healthy functioning of the human body and can lead to serious health issues. Current, prevalent tests can be expensive, cumbersome and not conducive to long-term monitoring of sleep. However, the emergence of the Internet of Things (IoT) paradigm offers an inexpensive alternative and can enable long-term, continuous monitoring of sleep. In this work, we tackle the problem of continuous sleep state segmentation in streaming data to temporally localize and recognize the sleep state. We show that predictive learning-based self-supervised representation learning can be used to learn robust features for both supervised and unsupervised sleep state segmentation. We obtain best accuracy of 96.2%(1 patient), an average accuracy of 81.8%(10 patients) on the wake vs sleep classification and best accuracy of 78.9%(1 patient), an average accuracy of 67.3%(10 patients) on the much harder multi-stage sleep classification task. We implemented the proposed framework on PYNQ board and with low power consumption of 3.27W and latency of 1.75s which offers a way forward for real-time, continuous sleep state segmentation.