Study Objective: Actigraphy is widely used in sleep studies but lacks a universal unsupervised algorithm for sleep/wake identification. This study develops an automated algorithm to effectively infer sleep/wake states. Methods: We propose a Hidden Markov Model (HMM) based unsupervised algorithm that automatically categorizes epochs into sleep/wake states. To evaluate the performance, we applied our algorithm to an Actiwatch dataset collected from 82 2-year-old toddlers, where epoch-byepoch comparisons were made between our algorithm and that in the Actiwatch software. Results: HMM identified more sleep epochs (earlier sleep onset and later wake-up) compared to the Actiwatch software for 386 (87.5%) and 423 (95.9%) out of 445 days for sleep start and end respectively. For the discrepant sleep epochs, 47.5% were zeros and the mean activity count per epoch was 33.0 (SD=29.5), suggesting immobility. HMM identified more wake epochs at sleep start for 21 days (4.8%), and 9.6% of the discrepant wake epochs were zeros and the mean activity count per epoch was 193.3 (SD=166.0), suggesting active epochs. The estimated HMM parameters can differentiate relatively active and sedentary individuals. A parameter denoted as σ for the wake state represents the variability in activity counts, and individuals with higher estimated σ values tend to show more frequent sedentary behavior. Conclusions: Our unsupervised data-driven algorithm overcomes the limitations of current ad hoc methods that often involve variable selection, threshold setting, and model training steps. Real data analysis suggests that it outperforms the Actiwatch software. In addition, the estimated HMM parameters can capture individual heterogeneities in activity patterns that can be utilized for further analysis.Significance: Current sleep/wake identification algorithms for actigraphy are often labor-intensive in model training steps, subjective in variable selection or threshold setting, and ad-hoc in the limited use of each trained algorithm in one dataset. Our proposed Hidden Markov Model-based algorithm is unsupervised that saves manual work and is also directly applicable to data from different sources. The unsupervised algorithm expands the application of actigraphy in large epidemiologic studies as well as in cases where intrusive polysomnography is hard to use, such as in pediatric populations. As an added benefit, the estimated HMM parameters can capture individual variabilities in sleep and activity patterns and one can use the information for further analysis.* Hidden Markov Model, HMM; Standard deviation, SD.

Physical activity, screen viewing, sleep, and homework among children have been independently linked to health outcomes. However, few studies have assessed the independent associations between time spent in daily activities and children’s physical and mental health. This study describes time spent in four activities among primary school students in Shanghai, and examines the relationship between daily time-use patterns and obesity and mental health. The representative sample consists of 17,318 children aged 6–11 years in Shanghai. Time spent in moderate to vigorous physical activities (MVPA), screen viewing, sleep, and homework was measured by validated questionnaires. Logistic regressions were performed. We also fitted generalized additive models (GAM) and performed two-objective optimization to minimize the probability of poor mental health and obesity. In 2014, 33.7% of children spent ˂1 hour/day on MVPA, 15.6% spent ≥ 2 hours/day on screen viewing, 12.4% spent ˂ 9 hours/day on sleep, and 27.2% spent ≥ 2 hours/day on homework. The optimization results suggest that considering the 24-hour time limit, children face trade-offs when allocating time. A priority should be given to the duration of sleep and MVPA. Screen exposure should be minimized to save more time for sleep and other beneficial activities.

Wearable devices have been increasingly used in research to provide continuous physical activity monitoring, but how to effectively extract features remains challenging for researchers. To analyze the generated actigraphy data in large-scale population studies, we developed computationally efficient methods to derive sleep and activity features through a Hidden Markov Model-based sleep/wake identification algorithm, and circadian rhythm features through a Penalized Multi-band Learning approach adapted from machine learning. Unsupervised feature extraction is useful when labeled data are unavailable, especially in large-scale population studies. We applied these two methods to the UK Biobank wearable device data and used the derived sleep and circadian features as phenotypes in genome-wide association studies. We identified 53 genetic loci with p<5×10 −8 including genes known to be associated with sleep disorders and circadian rhythms as well as novel loci associated with Body Mass Index, mental diseases and neurological disorders, which suggest shared genetic factors of sleep and circadian rhythms with physical and mental health. Further cross-tissue enrichment analysis highlights the important role of the central nervous system and the shared genetic architecture with metabolism-related traits and the metabolic system. Our study demonstrates the effectiveness of our unsupervised methods for wearable device data when additional training data cannot be easily acquired, and our study further expands the application of wearable devices in population studies and genetic studies to provide novel biological insights.