This study examined physical activity patterns among at-risk individuals using wrist-worn wearables. Step counts, distances, and energy expenditure (EE) were measured over 1 week in a real-life setting between November 22, 2021, and December 2, 2021. Participants aged >30 years who had risk factors based on metabolic syndrome diagnostic criteria and were currently using wrist-worn devices (eg, smartwatches and smart bands) were included. The risk factors included blood pressure ≥130/85 mm Hg, fasting blood sugar ≥100 mg/dL, triglyceride levels ≥150 mg/dL, high-density lipoprotein level <40 or <50 mg/dL (for male and female individuals, respectively), and waist circumference ≥90 or ≥85 cm (for male and female individuals, respectively). The number of study participants was selected based on the analysis results using G*Power (Heinrich Heine Universität Düsseldorf) and previous similar research cases. First, we used G*Power to perform an ANOVA because the data of the study participants would be measured repeatedly. The effect size (Cohen f) was set to be 0.25, and the significance level (α) and power (1 – β) were assumed to be 5% and 95%, respectively. As a result, a total sample size of 36 was calculated. Next, we considered the previous work by Huh et al [26], which applied wearable technology to patients with metabolic syndrome and recruited a total of 53 people. However, 33 patients dropped out during the 12-week study period due to the withdrawal of consent, device malfunction, and the loss of follow-up. We finally decided to recruit 60 study participants to achieve a sufficient effect. We used the Seoul National University mailing system to recruit research participants. Starting on November 3, 2021, we sent 2 emails to all members of the university; this lasted until November 19, 2021, when the recruitment was completed. The purpose of the study, eligibility criteria for study participants, and research procedures were provided via email. A detailed explanation of the study was provided in a web-based meeting after all written informed consent was obtained. Before physical activity measurements, all participants completed a web-based questionnaire through Google Forms, which collected personal health data and details about their physical activity (including type, intensity, and duration). The participants wore their own wrist-worn wearables for ≥10 hours per day with the physical activity measurement function activated for 1 week.

Following the 1-week period of physical activity measurements, the participants were asked to complete another web-based questionnaire to assess user experience, including cognition, context, applicability, and behavioral changes. Of the 60 participants initially included in the study, 13 were excluded due to missing or limited baseline data (n=3) or the unavailability of physical activity data from the database (n=10). Consequently, the analysis involved 47 participants who met the inclusion criteria and had data available for analysis.

During the 7-day measurement period, the participants activated the measurement function on their wrist-worn wearables, generating a total of 329 person-days of data. The study included wrist-worn wearables from the Apple Watch, Samsung Galaxy Watch, and Xiaomi Mi Band series (Multimedia Appendix 1). Data were collected continuously throughout the measurement period, with each device automatically storing individual data and synchronizing it with the participants’ cell phones.

Upon the completion of the measurement period, each participant exported their individual data through the official data export system and submitted the data via email. We provided the participants with detailed instructions specific to the manufacturer, version, and brand of their wrist-worn wearables. We then decompressed the data files and preprocessed the data to extract the selected variables, including step count, distance, EE, and duration with the start and end points. These variables were then merged into a unified data format for analysis.

The time-series clustering method was applied in 3 steps to identify clusters representing the physical activity patterns of the study participants. First, due to variations in data recording formats among the different wrist-worn device brands, the data were edited to ensure consistency. Data from different brands were standardized into the same format. For instance, the Samsung Galaxy Watch series record EE as “kcal per minute” during wearing, whereas the Apple Watch and Xiaomi Mi Band series record EE as “cal” and “kcal,” respectively, for each distinct activity. We converted these data into 10-minute “kcal” EE values. For activities performed for >10 minutes while wearing the Apple Watch or Xiaomi Mi Band series, the activity duration was divided into 10-minute intervals, assuming that a consistent amount of energy was expended during each interval.

Second, the data were divided into weekdays (Monday to Friday; 235 person-days) and weekends (Saturday and Sunday; 94 person-days), with the clustering method applied separately to each group. We calculated the average EE for weekdays and weekends based on the daily 10-minute EE values. It was assumed that the EE was 0 when the participant was not wearing the device. Figure 1 illustrates an example of physical activity measurements for 7 person-days, with each data point representing the EE in kcal per 10 minutes over the course of 1 week.

Third, time-series clustering was conducted through TADPole clustering, a recently developed technique that allows faster clustering by implementing a cutoff value to determine clusters [24,25]. To determine the optimal clustering model, we analyzed the expected number of clusters (parameter k) and the cutoff value. For the third step, we used the dtwclust package in RStudio [27]. This package offers a range of functions for conducting time-series clustering, including the TADPole clustering method. We had to specify certain parameters, including the cluster type, the number of clusters, the cutoff value, and the window size. Since we chose TADPole as the cluster type, we did not need to specify the distance parameter. As for the cutoff value and window size, we adjusted them based on the volume of data. Given that our study used 144 data points, we selected values that fell below this threshold. As the optimal number of clusters and cutoff values were unknown, cluster evaluation was performed using the silhouette index, which is a popular cluster validity index. Based on the cluster evaluation, the model with the highest silhouette index was selected as the optimal clustering model.

The demographic characteristics of the participants, including sex, age, work type (sitting, standing, etc), daily EE (weekdays and weekends), physical activity changes after using wrist-worn wearables, weekly physical activity patterns, and number of risk factors (1 or >1), were recorded. For categorical variables, the number and proportion for each category were presented, along with the P value calculated using the χ² or Fisher exact test for variables with counts <5. For numerical variables, mean and SD with P values were calculated using the 2-tailed t test.

The association between weekly physical activity patterns and participant characteristics was analyzed using a logistic regression model. The regression model was evaluated in terms of pseudo-R², accuracy, Hosmer-Lemeshow goodness of fit, and the receiver operating characteristic curve. The results of the logistic regression were presented as odds ratios with 95% CIs and the corresponding P values. Statistical significance was taken as P≤.05. The statistical analyses were performed using RStudio (version 2022.07.2+576; Posit) [28].

The study was approved by the Institutional Review Board of Mokpo National University (approval MNUIRB-210625-SB-014-01). All participants provided informed consent before study participation. The submitted data were anonymized before analysis. Participants who provided data and finished the web-based survey received a compensation of ₩100,000 (US $77.65).

Among the 47 participants, 23 (49%) were male and 24 (51%) were female (Table 1). In terms of age, 30 (64%) participants were aged <40 years, whereas 17 (36%) were aged ≥40 years. The majority (n=42, 89%) of the participants had a sedentary job. The average EE during weekdays was 223 (SD 175) kcal, whereas that on weekends was 191 (SD 164) kcal. After using wrist-worn wearables, 29 (62%) participants reported a decrease or no change in physical activity, whereas 18 (38%) participants reported an increase. In terms of health risk factors, 25 (53%) participants had only 1 risk factor, whereas 22 (47%) had >1 risk factors.

The time-series cluster analysis resulted in the highest silhouette index when there were 2 clusters (k=2) for both weekdays and weekends. Therefore, 2 clusters each were distinguished for weekdays and weekends (Figure 2). The left and right columns of Figure 2 represent the weekday and weekend clusters, respectively. Each cluster included data from at least 2 participants, and 2 distinct cluster types were distinguished with different starting times for physical activity. The “early bird” type (represented by blue dots in Figure 2) initiated physical activity after 6 AM, whereas the “night owl” type (represented by orange dots in Figure 2) began physical activity before 6 AM.

Among the 37 early birds on weekdays, 57% (n=21) were female and 76% (n=28) were in their 30s, with female participants in their 30s accounting for the highest proportion (n=18, 48%). Among the 10 night owls on weekdays, 70% (n=7) were male and 80% (n=8) were in their aged ≥40 years, and 40% (n=4) were male individuals aged ≥40 years. Weekend physical activity patterns were mostly from early birds (n=45, 96%) regardless of sex or age. Out of the 45 early birds on weekends, 51% (n=23) were female and 64% (n=29) were in their 30s, and 40% (n=18) were female individuals aged <40 years. There was 1 participant per sex and age group who was a night owl on weekends, and there was no one who was a night owl on both weekdays and weekends.

Figure 2A shows the physical activity patterns of the 37 (79%) out of 47 participants belonging to cluster 1 (early birds) on weekdays. Physical activity occurred between 6 AM and 8 PM, with most activities being <10 kcal (mean 2.98 kcal). Meanwhile, cluster 2 included 10 (21%) “night owls” on weekdays (Figure 2B). Physical activity for cluster 2 typically started at midnight and ended before 4 PM. Cluster 2 also exhibited greater EE, with an average of 9.51 kcal.

During the weekends, the majority (45/47, 96%) of participants were early birds (cluster 3; Figure 2C). Cluster 3 had an average EE of 2.94 kcal. Cluster 4 (night owls) included only 2 (4%) participants on weekends (Figure 2D) and demonstrated a higher EE (average 8.60 kcal).

Although the average EE was similar between early birds and night owls, their physical activity patterns differed on weekdays and weekends. Regardless of the cluster type, physical activity tended to be shorter in duration on weekdays (Figure 2A and B), becoming longer and more continuous on weekends. Cluster 3, representing early birds on weekends, exhibited up to 10 consecutive physical activity periods, which is equivalent to 100 minutes (4-8 AM; Figure 2C). Cluster 4, representing the night owls on weekends, had the highest total EE of 507.59 kcal and up to 16 consecutive physical activity periods, which is equivalent to 160 minutes (4-8 AM; Figure 2D).

Based on the analysis of physical activity patterns, 2 groups were identified: the stable group and the shifting group. The stable group included individuals who maintained the same physical activity pattern on weekdays and weekends, regardless of whether they were classified as early birds or night owls. Among the 47 participants, 35 (74%) belonged to the stable group, exhibiting early bird physical activity patterns consistently throughout the week. There were no participants belonging to both clusters 2 and 4. On the other hand, the shifting group included individuals whose physical activity patterns differed between weekdays and weekends. There were 12 (26%) participants in the shifting group; 10 (21%) participants displayed an early bird pattern during weekdays (cluster 1) but changed to the night owl pattern on weekends (cluster 4). The remaining 2 (4%) participants exhibited the opposite pattern, that is, a night owl pattern during weekdays (cluster 4) and an early bird pattern on weekends (cluster 1).

Demographic descriptive statistics for the physical activity pattern groups, including the results of the χ² test for categorical variables and the t test for continuous variables, are presented in Table 2. There were no significant differences between the physical activity pattern groups except in age (P=.001) and EE (P<.001 for weekdays; P=.003 for weekends).

A logistic regression model was used to examine the associations of sex, age, and the number of health risk factors with weekly physical activity patterns (Table 3). Logistic regression model accuracy and diagnostic results are presented in Multimedia Appendix 1. Sex (P=.45) and the number of health risk factors (P=.33) were not significantly associated with the physical activity pattern. In contrast, age showed a statistically significant association with physical activity patterns; the higher age group had higher odds of differences between weekday and weekend physical activity patterns (odds ratio 8.68, 95% CI 1.95-48.85; P=.007).

To account for the possibility of reverse causality, we conducted another logistic regression analysis with the number of health risk factors as the outcome variable. Despite this adjustment, there were no significant associations between physical activity patterns and health risk factors (P>.99; Multimedia Appendix 1).

In this study, we assessed the effectiveness of the TADPole clustering method for identifying physical activity patterns from wearable data. We also explored the association between these patterns and health risk factors. We found that physical activity patterns on weekdays and weekends were categorized as either daytime (early bird) or nighttime (night owl) patterns. Furthermore, 2 groups were distinguished: 1 with consistent physical activity patterns on weekdays and weekends (stable group) and the other with different patterns between weekdays and weekends (shifting group). Age significantly influenced physical activity patterns.

We found that physical activity patterns on weekdays and weekends differed as age increased. Our findings shed light on previously unaddressed or overlooked associations between physical activity patterns and health risk factors. Many previous studies did not report the association between age and physical activity patterns [17,18], and this association was reported only in a few studies [29-32]. Some of these studies reported results consistent with our findings. A study by Caspersen et al [29], which analyzed physical activity patterns according to sex and age, found that not only inactivity but also vigorous activity increased with age. In their study, those aged 18-29 years showed a pattern with the lowest vigorous activity and the highest sustained physical activity, whereas those aged ≥75 years showed a pattern with the highest vigorous activity and the lowest sustained physical activity, indicating differences in physical activity patterns by age. Another study by Rossen et al [32], which analyzed the physical activity patterns of individuals with diabetes for 2 years, found that the younger the age, the more physical activity increased 2 years later, showing that physical activity patterns can vary depending on age.

The physical activity patterns identified in our study were similar to chronotypes, which categorize individuals into morning type (M-type), evening type (E-type), and intermediate type (N-type) based on their preferred timings of activities and sleep [33]. M-type individuals are typically early birds, whereas E-type individuals are night owls who prefer late activity and sleep schedules. N-type individuals do not fall strictly into either category, and most adults belong to this type [33]. Previous studies have suggested that E-type individuals have lower physical activity levels and a higher risk of metabolic syndrome [34,35]. However, this study found that the individuals with night owl tendencies, that is, E-type individuals, exhibited higher physical activity levels compared to early birds, that is, M-type individuals. It is important to note that most participants in our study reported engaging in sedentary work, indicating that the increased physical activity levels among night owls were likely due to leisure activities rather than occupational tasks. This suggests that physical activity patterns are not determined solely by chronotype and that other factors, such as health awareness, can have a significant impact. It is worth noting that an individual’s chronotype and activity times may vary based on age and occupation, potentially leading to health issues if not addressed [36]. Therefore, it is crucial to make efforts to achieve the recommended level of physical activity regardless of the specific activity pattern or chronotype.

None of the participants in our study met the criteria for “weekend warriors,” which refers to individuals engaging in 1 or 2 sessions of physical activity, particularly on weekends, per week, consuming at least 1000 kcal [37]. Although the combined EE for cluster 2 (weekday night owls) and cluster 4 (weekend night owls) was the closest to that of weekend warriors at 735 kcal, no participants were included in both clusters. In a study conducted by Jang et al [38] in South Korea, only 2.1% of the participants were classified as weekend warriors, but there was no significant difference in metabolic risk between weekend warriors and the regularly active group. The weekend warrior physical activity pattern, which is popular in the United Kingdom, the United States, and Latin America [39], is associated with several health benefits, including a lower risk of obesity [38] and all-cause mortality [37]. A study of Chinese adults found that the weekend warrior physical activity pattern was associated with a lower risk of metabolic syndrome, hypertension, and diabetes in both male and female individuals [40]. Promoting physical activity guidelines may increase the number of weekend warriors in South Korea, where sedentary jobs are common (Table 1).

The TADPole clustering method addresses the limitations of previously popular time-series clustering techniques such as the k-means and hierarchical clustering methods in data analysis. As we mentioned in the introduction, the k-means clustering method often produced unstable results with clusters changing each time, whereas the TADPole clustering method consistently provided reliable clustering outcomes. However, similar to k-means, we iteratively performed the analysis to obtain the optimal clustering results. The TADPole clustering method stands out by providing more reliable results compared to hierarchical clustering and DBSCAN, as it uses all available data without any loss or exclusion. Additionally, our study demonstrated the feasibility of the TADPole clustering method, showing its suitability for handling large and multidimensional data such as wearable data.

A key advantage of our study was that we used multiple wrist-worn device models. Although wearables provide data in different formats depending on the model, we standardized the data into a single format to successfully conduct statistical analyses. Furthermore, the TADPole clustering method allowed us to overcome the limitations of hierarchical and k-means clustering, which are commonly used time-series clustering methods, resulting in robust and reliable findings.

This study had several limitations. It only included 47 participants, which may not have been sufficient to generate meaningful results. The measurement period for physical activity was only 7 days, which may not have been representative of the daily physical activity. Additionally, we assumed that the EE for physical activities exceeding 10 minutes was consistent across 10-minute intervals. Although these assumptions may not perfectly reflect reality, they were considered reasonable given that the participants were going about their normal daily routines.

This study successfully performed time-series clustering using various wrist-worn device models and found TADPole clustering to be a suitable tool for analyzing the data. Physical activity patterns on weekdays and weekends could be categorized into “early birds” and “night owls,” and these patterns were significantly influenced by age. To address the limitations of our study, additional studies with larger sample sizes are required.