Quantification of Physical Activity and Sleep Behaviors with Wearable Sensors: Analysis of a large-scale real-world heart rate variability dataset

Wearable monitoring devices, such as smartwatches, are used for monitoring personal health, fitness, health behaviors and well-being in daily life. Nowadays, wearable devices are popular and many consumers use them, in particular, to record their physical activity and sleep. Data recorded with wearable devices is an example of real-world data that can provide practical observations and insights on health and wellness, but its analyses pose challenges for research. Consumers conduct continuous recordings with wearable devices in non-research settings. Hence, any analysis of wearable real-world monitoring data must take into account the limitations and inaccuracies of the data, as well as sampling biases and incomplete representativeness of the population that arise from the uncontrolled data collection setting. To date, there are no well-established methods for analyzing health behaviors and well-being from continuous wearable monitoring data. Consequently, real-world health monitoring data is not commonly used for research although it could provide valuable observations and insights on health behaviors and well-being. This thesis work aims at analyzing a large-scale real-world dataset of wearable heart rate variability (HRV) recordings to quantify the behaviors of physical activity (PA) and sleep that are one of the most important health behaviors. Specifically, the thesis focuses on the quantification methods and temporal patterns of PA behavior, as well as the associations that PA, alcohol intake and other lifestyles have with sleep. In addition, this thesis work aims to evaluate the feasibility to use real-world wearable monitoring data with applicable analysis methodologies for scientific research, and to demonstrate the observations and data-driven hypotheses that the results provide. The study material was an anonymized real-world HRV monitoring dataset of 52,273 Finnish employees, which was gathered and prepared by Firstbeat Technologies Oy (Jyväskylä, Finland), a Finnish company providing and developing HRV analytics for stress, recovery and exercise. The dataset included three-day continuous HRV recordings performed in free-living settings combined with self- reports of alcohol intake, work and sleep times. The recordings were originally performed for a routine wellness program (Firstbeat Lifestyle Assessment) provided for the employees by their employers as a part of preventive occupational healthcare and health promotion program. For the analysis of this thesis, PA behavior was quantified from the recordings using an HRV-based estimate of the oxygen uptake. Sleep was quantified by the regulation of the autonomic nervous system (ANS) using traditional HRV parameters and novel HRV-based indices of recovery. Both statistical and machine- learning methods were employed in the analysis for the thesis results. Temporal variations in PA behavior were observed: the amount of PA was highest at the weekends and at the beginning of the year. The amount of PA quantified by the absolute oxygen consumption was higher for men than for women, and higher for younger than older subjects, and also higher for individuals of normal weight than obese. However, PA levels were more similar between the subjects when their physical fitness level was considered in quantifying PA. Moreover, PA behavior was associated with sleep. After a day including PA, the parasympathetic regulation of the ANS and recovery during sleep were diminished, but regular PA seemed to increase parasympathetic regulation of the ANS and aid recovery during sleep. The most important predictor for ANS regulation during sleep was, however, acute alcohol intake. Acute alcohol intake dose-dependently diminished the parasympathetic regulation of the ANS and recovery during sleep, an effect that was already observable after only 1–2 standardized units of alcohol. Moreover, the same alcohol intake, normalized by the body weight, seemed to affect the ANS regulation more in younger subjects than in the older ones, but was similar for both sedentary and physically active subjects, as well as for both men and women. Many of the results obtained in this thesis accord with the findings of previous studies, such as the higher PA level on weekends, the higher amount of absolute intensity PA in men, younger and normal weight subjects, and the relationship of PA and alcohol intake with the ANS regulation during sleep. On the other hand, the results of this thesis provide new observations, for example, about the interaction between alcohol intake and subject’s background characteristics that could not have been studied before due to the limited and homogenous study populations. In conclusion, the results of this thesis demonstrates that real-world wearable monitoring data can be feasible for scientific research and its results not only supports the findings of existing studies but also provides new observations, insights and data-driven hypotheses. The real-world evidence facilitates our understanding of aspects of health behaviors and wellness that cannot be studied in the more traditional, controlled research settings. These real-world insights can be further used for designing more personalized and targeted health interventions and as tools for promoting health and well-being.