Accurately Inferring Physical Activity Levels and Sleep From Wrist-Worn Actigraphy Recordings With Sample Rates as Low as 10 Hz

Inferring longitudinal Physical Activity (PA) levels and sleep timings from wrist-worn sensors may facilitate personalized insights into day-to-day profile assessments and can be used to monitor a range of physical- and mental-health outcomes, including towards symptom monitoring and rehabilitation. We used the publicly available CAPTURE-24 dataset, comprising 148 participants with ~24-hour concurrent three-dimensional wrist-worn accelerometer data and minute-by-minute labels used as ground truth: sleep, sedentary, light, moderate-vigorous PA. First, we down-sampled the raw accelerometry data to 10 Hz to ensure the generalizability of our methodology across longitudinal studies which typically use similarly low sample rates for actigraphy. Subsequently, we computed four complementary acceleration summary measures and 10 additional smoothened outputs for each acceleration summary measure to derive 44 features characterizing minute-by-minute PA. These features were presented into different classifiers casted as a 4-class classification problem. We trained the model using the first 98 participants and assessed model performance and generalization on the remaining 50 participants. Using a random forest classifier, we demonstrated accurately estimating PA levels and sleep with 87% overall accuracy (F1-score=0.80) including 98.6% correct sleep detection. These findings processing the down-sampled actigraphy data to 10 Hz match or exceed state-of-art results recently reported in the literature achieved using considerably more sophisticated and time-consuming methods (including deep learning) which required actigraphy data sampled at 100 Hz. Collectively, these findings support the deployment of longitudinal, large-scale actigraphy data with sample rates as low as 10 Hz, towards accurately estimating personalized day-to-day PA and sleep profiles in healthcare community studies.