JMIR Publications

ABSTRACT

Background:

The proliferation of wearable devices collecting activity and heart rate data has facilitated new ways to measure sleep-wake duration unobtrusively and longitudinally. Most existing sleep-wake identification algorithms are based on activity only and trained on expensive and laborious Polysomnography (PSG) annotations. Heart rate can also be reflective of sleep-wake transitions, which motivates its infusion in an unsupervised algorithm. Moreover, a personalized approach remains to be developed to deal with inter-subject variance in their sleep-wake patterns.

Objective:

We aim to develop a personalized sleep/wake identification approach using multi-faceted data in an unsupervised way, to explore the benefits of incorporating heart rate in the algorithm, and to compare it with an existing commercial wearable device, Fitbit Alta.

Methods:

In this study, a total of 14 community-dwelling elderly wore wearable devices (Fitbit Alta) 24/7 over 3 months and their heart rate and activity data were collected. After preprocessing, we built a personalized model for each participant to distinguish sleep/wake states based on the individual’s data. We proposed the use of hidden Markov models (HMM) and compared different modelling schemes. With the best model selected, sleep-wake patterns were characterized by estimated parameters in HMM and sleep/wake states were identified.

Results:

In our study, a total of 927 days of data were collected and analyzed for 14 participants. After the implementation of our approach, the estimated parameters in HMM can reflect the inter-subject variability in heart rate and activity intensity during hidden sleep and wake states. In a case study, the fusion of heart rate and activity data in our approach help detect 14.36% more wake epochs comparing to approach with single-source data. Our algorithms agreed with Fitbit’s scoring results 86.70% on average over participants. It potentially identified more wake epochs at night comparing to Fitbit, suggesting a better performance.

Conclusions:

Our approach can be effectively implemented based on individual’s multi-faceted sleep-related data from a commercial wearable device in an unsupervised way. Personalized model is shown necessary considering the inter-subject variability in estimated parameters.