For the first time in a nationally representative adult population, we applied a data-driven clustering approach to identify population segments based on 24-hour physical activity data collected using a wearable device. Based on inputs of twenty-four hourly PA values, we found five distinct cluster profiles, which we describe as follows. Cluster 1 represents a “High PA” group with escalated daytime PA. Cluster 2 portrays a “Low PA” group with a small increase in PA. Clusters 3 and 4 display acceleration in rest/sleep time, an indicator of circadian rest-activity rhythm (CR) disturbance43. Specifically, Cluster 3 refers to a “Mild CR disruption” group that exhibits reduced activity during the day and increased nocturnal activity. Cluster 4 shows a substantially high rest/sleep hour PA, referred to as “Extreme CR disruption”. Lastly, Cluster 5 represents “Very low PA”.
We demonstrated that clusters are significantly associated with demographic characteristics, inflammatory biomarker levels, all-cause mortality, as well as biological age. Across all health-related outcomes, “High PA” group (Cluster 1) tends to have the best performance, with the lowest values of inflammation levels, mortality, and biological age advance. This was followed by “Low PA” (Cluster 2), “Mild CR disruption” (Cluster 3), and “Extreme CR disruption” (Cluster 4). “Very low PA” (Cluster 5) group performed worst, with the highest inflammation levels, biological age advance, and mortality risk (see Fig. 5).
There were, however, a few exceptions. “Extreme CR disruption”, consisting of young adults aged 30–40 years, was significantly associated with increased inflammatory biomarkers and accelerated biological age, but not with all-cause mortality and medical histories. This finding suggests that young adults with circadian misalignment may seem ostensibly healthy because they have no apparent signs of medical conditions and show high levels of physical activity, but in fact, are undergoing health deterioration and unhealthy aging. In middle-aged adults, having some degree of circadian cycle disturbance together with a low PA level (“Mild CR disruption”) resulted in substantially higher inflammatory biomarker levels, mortality risk, and biological age compared to having low PA alone (“Low PA”). This highlights the growing importance of circadian alignment in older populations to achieve healthy longevity.
Unlike physical activity or nutrition, there is still a lack of understanding regarding how to utilize or correct biological timing for health benefits44,45. Current public health interventions are largely focusing on increasing physical activity levels or eating healthy, with less attention on targeting the clock. With growing evidence demonstrating the significance of circadian disruption on performance, well-being, and longevity, clock-based therapeutics have enormous potential for maximizing health benefits and promoting healthy aging40, 45–47. In particular, the increasing uptake of digital devices (e.g., smartphones and wearables) facilitates the collection of accelerometry and other behavioral data over a larger time at greatly reduced costs from a large population5. Coupled with machine learning algorithms, digitization of such passive behavior data has an unrecognized potential as novel digital biomarkers for longevity and advancing personalized interventions, automated health event prediction, and population-level prevention. As an implication of this study, we can utilize wearable data as a digital biomarker and deliver personalized intervention via digital devices to successfully migrate “unhealthy or at-risk” individuals to “healthy” clusters. Young adults with impaired circadian cycle, for example, will be instructed to shift sleep and wake times to earlier hours. Older adults with low activity levels, on the other hand, will be recommended to be more physically active.
There are potential limitations of this study. First, the validity of the feature selection must be verified on new data, unseen from the model during the development phase. Second, this is a retrospective analysis and cannot establish causal relationships between the observed associations. Third, we use only 7-day accelerometer data, and a longer duration of monitoring would provide a more precise and accurate classification of clusters. Fourth, unmeasured environmental factors or residual confounding could have affected accelerometry measurements. Lastly, non-wear time and missing accelerometry measures may influence the activity output. However, the impact is minimal as we selected participants with complete 5-minute epoch information in the analysis.
Nevertheless, this study offers the following contributions over previous research. This study is the first to use wearable-based physical activity data to segment a nationally representative sample of the U.S. population. A novel, detailed resolution of 24-hour activity profile elucidates distinct cluster profiles and highlights circadian misalignment and rhythm disruption to play a critical role in longevity measures of inflammation, biological age, and mortality. With this work, we add a meaningful contribution to current research in the field demonstrating the potential for the digitization of human longevity measures based on wearable-based physical activity. A digital biomarker for longevity has enormous potential for digital phenotyping, personalized intervention, population-level prevention, and remote monitoring of people’s health. It is also a critical step toward achieving the aim of precision medicine. Future studies with prospective and repeated assessments using digital devices are warranted.