A Two-stage HMM Model for Sleep/Wake Identification via Commercial Wearable Device

Good sleep habit is essential to maintain a quality life. Sensor-based wearable devices have been increasingly deployed to monitor activity and measure sleep duration in a non-intrusive, affordable, and portable way. Existing algorithms for detecting sleep and wake in wrist-worn devices are mainly based on activity count or accelerometer data inference. It is validated that sleep can be reflected by various vital signs, including physical activity, heart rates, and pulse oximetry. However, little attention has been paid to heart rates measurements for sleep/wake identification in commercial wearable devices together with activity information. Our study developed an unsupervised and personalized algorithm to infer sleep and wake states using heart rates and step counts based on hidden Markov models. The fusion of two HMMs successfully dealt with multi-granularity data and predicted sleep/wake states in the minimal granularity. The proposed algorithm was illustrated through a real-life case study. The agreement between our algorithm and Fitbit's scoring was 89.35%. The proposed algorithm enabled identifying more afternoon naps, earlier sleep onset compared to Fitbit's scoring. The results showed that heart rates were informative when distinguishing sleep and wake while compensating estimations driven from step counts.