In an aging society, early detection and prevention of frailty are urgent priorities. Frailty is a multifaceted concept encompassing physical, cognitive, and social aspects, all of which play critical roles in comprehensively assessing the health status of older adults [1,2]. Among these, social frailty precedes physical frailty and has been reported to promote unhealthy aging while increasing the risks of all-cause mortality and functional impairment [3,4]. Bunt et al [5] proposed that, in addition to general resource–related factors such as living alone and economic difficulties, social resource factors such as connections with neighbors and friends, as well as basic social needs such as social behavior, social isolation, and social support, are critical components of social frailty. Social frailty profoundly affects the lifestyles of older adults and has been associated with a poor nutritional status, depression, reduced physical capability, and decreased activity levels [6]. Furthermore, it has been linked to declines in activities of daily living [7] and instrumental activities of daily living [8,9]. Rest-activity rhythm (RAR) has long garnered attention as an established indicator for evaluating health conditions related to such lifestyle factors.

Previous studies using actigraphy have used RAR parameters derived from activity trackers to predict various health outcomes, including physical frailty, cognitive decline, and depression [10-14]. These parameters quantify 24-hour cycles of rest and activity based on activity counts and have been widely used in health research. However, their application to social frailty remains underexplored [15], and studies leveraging wearable sensors for this purpose are particularly scarce.

Recently, wearable devices such as Fitbit smartwatches have enabled the collection of detailed physiological and behavioral data, including heart rate, respiratory rate, and sleep data in addition to traditional activity data. These advancements have spurred the development of innovative digital biomarkers such as heart rate–based rhythm metrics, which complement conventional activity count–based indicators [16,17]. Fitbit devices, known for their affordability and accessibility, have demonstrated accuracy comparable to research-grade devices for measuring step count and heart rate [18-20]. Despite their potential, the use of wearable devices to evaluate social frailty and its associated risk factors remains limited, highlighting the need for further investigation, such as studies similar to the one reported in this paper.

Fitbit and other wearable devices provide step count, heart rate, and other data that allow for the calculation of digital biomarkers to predict health outcomes. However, studies using raw data from these devices to compute novel metrics and evaluate their associations with health outcomes remain insufficient.

The primary aim of this study was to identify digital biomarkers associated with social frailty. Social frailty has been linked to reduced daytime physical activity, as demonstrated in previous research [7,9]. Although there is limited evidence directly connecting social frailty to heart rate, it is plausible that decreased physical activity in individuals with social frailty leads to lower overall heart rate variability (HRV) and altered heart rate patterns. Therefore, we hypothesized that the frailty group would exhibit significant associations with RAR parameters (nonparametric and extended cosinor analysis) and heart rate–based parameters (eg, nighttime heart rate and resting heart rate [RHR]), assessed by leveraging Fitbit data. These parameters have the potential to shed light on the physiological and behavioral characteristics of social frailty.

The secondary aim of this study was to explore the relationship between RAR metrics derived from Fitbit data and cognitive function. While previous research has linked nonparametric and cosinor-based RAR metrics to cognitive decline, most studies have relied on research-grade actigraphy devices. This study introduces novel RAR metrics derived from Fitbit data to exploratorily assess their associations with cognitive function.

This study was designed as a cross-sectional analysis to investigate the relationship between digital biomarkers and social frailty in community-dwelling older adults in Akita Prefecture, Japan. To determine the appropriate sample size for the logistic regression analysis, we used G*Power (version 3.1.9.6) with the likelihood-ratio test as the test statistic. The estimation indicated that a minimum of 82 participants would be required to detect a clinically significant effect based on the following parameters: 3 groups, a significance level (α) of .05, a statistical power of 80%, and an odds ratio (OR) of 2.0 [21]. Considering a potential dropout rate of 20% to 30%, we recruited a total of 102 older adults between October 2023 and December 2024. Recruitment sites included municipally organized programs conducted at care management centers in the Yabase, Kawabe, Ushijima, Goshono, Iijima, and Shimoshinjo districts of the city of Akita, as well as in the city of Yurihonjō and the town of Mitane. Additional participants were recruited from a community-based exercise group regularly followed by the principal investigator held at a temple in the Katsuhira area. The inclusion criteria were (1) age of ≥65 years, (2) independent walking ability in daily life, and (3) residence at home. The exclusion criteria were (1) a central nervous system disorder, (2) requiring assistance in daily living, (3) a cardiac condition, or (4) requiring support or care as certified under Japan’s public long-term care insurance system.

Data collection used consumer-grade fitness trackers, specifically the Fitbit Inspire 2 and 3 (Google). The accuracy of Fitbit data has been evaluated in several studies. According to systematic reviews [22,23], Fitbit devices consistently demonstrate acceptable levels of accuracy for daily step counts and heart rate measurements despite a tendency to underestimate these metrics under certain conditions. Generally, Fitbit wearables provide activity assessment accuracy comparable to research-grade devices but tend to overestimate moderate to vigorous physical activity under free-living conditions. One of the major challenges faced by medical researchers when interpreting physical activity data is related to participants forgetting to wear the activity tracker [23]. However, the Inspire series of Fitbit devices, being waterproof and offering a battery life of up to 10 days, minimizes interruptions in data collection. These features contribute to ensuring high-quality data with fewer missing values, making them particularly advantageous for calculating rhythm metrics.

Participants were instructed to wear the device for at least 7 consecutive days. The research staff placed the device on the participants’ nondominant wrist, approximately 1 inch below the wrist bone, ensuring that the back of the device was in contact with the skin. Participants were instructed to wear the tracker at all times, including during bathing, whenever possible. However, they were advised not to use the device in situations prohibited by the manufacturer, such as diving under high water pressure or in high-temperature environments such as saunas.

This study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. The research protocol and informed consent forms were approved by the institutional review board of Akita University School of Medicine (approval 3063). All participants provided written informed consent before taking part. This study involved both the collection of wearable device data and in-person assessments, including evaluations of social frailty and cognitive performance. To ensure participant confidentiality and data protection, all collected data were anonymized before analysis. Participants did not receive monetary compensation, nor did they incur any costs by taking part in the study. As a token of appreciation, each participant received a booklet summarizing their individual results and interpretations.

Demographic data, including age (years), sex (female or male), and educational level (years), were collected from all participants. Social frailty was classified using the social frailty index developed by Makizako et al [24], which reflects the accumulation of social risk factors and indicates reduced social participation or roles. The index consists of five components: (1) living alone (yes), (2) talking with someone every day (no), (3) feeling helpful to friends or family (no), (4) going out less frequently compared with the previous year (yes), and (5) visiting friends sometimes (no). Participants with a total score of 0 were classified as robust, those with a score of 1 were classified as social prefrailty, and those with scores of 2 to 5 were classified as social frailty. Physical performance was assessed by measuring grip strength (kg) and usual walking speed (m/s) over a 5-m course. Cognitive performance was evaluated using 4 subtests from the National Center for Geriatrics and Gerontology–Functional Assessment Tool [25]. The word list memory (WLM) test included immediate recognition and delayed recall tasks, where the mean correct answers across 3 trials (range 0-10) and the number of correct answers in delayed recall (range 0-10) were summed to calculate a total score (range 0-20). The trail making test versions A and B was used to measure cognitive flexibility, with completion times recorded in seconds. Finally, the digit symbol substitution task was conducted to assess information processing speed, where the number of correct responses in 90 seconds was recorded.

After the monitoring period, devices were collected, and their stored data were synchronized via Bluetooth to a smartphone before being transferred to the Fitbit database. Researchers then accessed the Fitbit developer platform linked to a Google account [26], generated an access token (refreshed every 8 hours), and executed a Python script (Python Software Foundation) in Google Colaboratory to retrieve the raw step count and heart rate data at 1‑minute intervals. To ensure data integrity, periods without heart rate data were excluded, and corresponding step count data were marked as invalid; only time points with both valid heart rate and step count data were retained as clean data. Consistent with the work by Rykov et al [17], a valid day was defined as one with at least 20 hours of concurrent heart rate and step count data. On the basis of evidence that 5 days of wear yield reliable RAR parameters in older adults [27], only participants with at least 5 valid days were included in the study.

Raw data from Fitbit were used to extract digital biomarkers related to circadian rhythms and heart rate variation. Step count and heart rate data were processed using a modified version of the script developed by Rykov et al [17] to calculate both extended cosine-based and nonparametric circadian rhythm indexes.

The normality of the variables was assessed using the Shapiro-Wilk test. For variables following a normal distribution, multiple regression analysis was conducted with the WLM score as the dependent variable, and explanatory variables were selected based on Pearson correlation coefficients. For nonnormally distributed variables, the Spearman correlation analysis was conducted. The Spearman analysis was applied to the entire dataset (all participants) as well as separately to the robust, social prefrailty, and social frailty groups. Kruskal-Wallis tests, followed by Dunn post hoc tests, were used to evaluate differences across the 3 frailty groups. Finally, multinomial logistic regression models were constructed to identify predictors of social frailty status (robust, social prefrailty, and social frailty). On the basis of the results of the Kruskal-Wallis tests, models I and III included age- and sex-adjusted factors as independent variables, whereas models II and IV were unadjusted for these factors. All analyses were conducted using RStudio (version 4.4.1; Posit PBC).

The average device wearing time per participant was 186 (SD 26) hours, of which an average of 164 (SD 30) hours per participant consisted of complete data recorded on days with at least 20 hours of valid data. Of the 102 participants, only 86 (84.3%) had sufficient data for at least 5 days and were included in the analysis.

The participants had a mean age of 77.14 (SD 5.7; range 65-88) years, and most were women (78/86, 91%). The average daily step count was 7824 (SD 3794), and the mean heart rate was 72.3 (SD 5.9) beats per minute. According to the social frailty criteria defined by Makizako et al [24], the sample was categorized into robust (28/86, 33%), social prefrailty (39/86, 45%), and social frailty (19/86, 22%) groups.

Figure 2 presents the averaged 24-hour profiles of hourly heart rate and step count for all participants (n=86). Figure 3 shows the same 24-hour profiles stratified by social frailty status (robust, social prefrailty, and social frailty).

The demographic data for all participants and each social frailty group are summarized in Table 1, including key variables such as age, sex, and variables from physical and cognitive domains, as well as digital biomarkers. Only the key variables are shown in this table, with the full version available in Table S1 in Multimedia Appendix 2.

Figure 4 illustrates the results of the Spearman correlation analysis conducted across the entire dataset (all participants) and separately for the robust, social prefrailty, and social frailty groups. Age, physical domain, and cognitive domain variables are included in the analysis, whereas other variables are reported only if their correlation coefficients showed significant associations in at least 3 of the panels. Tables S2-S5 in Multimedia Appendix 2 provide detailed correlation coefficients.

Table 2 presents the results of the multiple linear regression analyses with WLM score as the dependent variable. In total, 2 models were analyzed. Model I included age and ICV × 100 as independent variables, whereas model II added one more variable to the predictors. ICV × 100 represents the step-based intrahourly CV scaled by multiplying the original values by 100 to enhance interpretability.

In model I, both age (β=−0.15, 95% CI −0.26 to −0.03; P=.01) and ICV × 100 (β=−0.04, 95% CI −0.07 to −0.01; P=.02) were significant predictors of WLM score. In model II, age (β=−0.13, 95% CI −0.25 to −0.02; P=.03) and ICV × 100 (β=−0.04, 95% CI −0.07 to 0.00; P=.03) remained significant.

Table 3 summarizes the results of the Kruskal-Wallis and Dunn post hoc tests for digital biomarkers across the robust, social prefrailty, and social frailty groups. The Dunn post hoc tests were conducted following the Kruskal-Wallis test to identify pairwise differences, and the P values from these tests were adjusted using the Bonferroni method. The table reports only the 17 variables with significant associations identified through the Dunn test (P<.05). Notably, comparisons between the robust and social frailty groups revealed significant differences in 13 biomarkers.

Table 4 summarizes the results of the multiple logistic regression analysis, with models I and III adjusted for age and sex and models II and IV excluding these adjustments. The robust group was used as the reference category.

In model II, the dRHR showed significant negative associations with both social prefrailty (OR 0.82, 95% CI 0.68-0.97; P=.02) and social frailty (OR 0.74, 95% CI 0.58-0.94; P=.02). Similarly, UpMesor.hr demonstrated significant associations with both social prefrailty (OR 0.57, 95% CI 0.46-0.71; P<.001) and social frailty (OR 1.45, 95% CI 1.09-1.94; P=.01).

In model IV, the α parameter.hr × 100 was significantly associated with social prefrailty (OR 1.05, 95% CI 1.01-1.09; P=.02). The SD of the heart rate showed a significant negative association with social prefrailty (OR 0.75, 95% CI 0.59-0.94; P=.02). For social frailty, ICV.st × 100 demonstrated a significant positive association (OR 1.04, 95% CI 1.01-1.08; P=.01).

The results of models I and III, which were adjusted for age and sex, were generally consistent with their corresponding models (model I corresponds to models II, and model III corresponds to model IV). The primary difference was observed in UpMesor.hr, which was significantly associated with both social prefrailty and social frailty in model II but only with social prefrailty in model I. Regarding other variables, dRHR, ICV.st × 100, the α parameter.hr × 100, and the SD of the heart rate showed consistent associations across corresponding models.

Our results lent support to the hypothesis that the frailty group would exhibit significant associations with RAR parameters (nonparametric and extended cosinor analysis) and heart rate–based parameters. The observed significant relationships with dRHR, UpMesor, and other heart rate–derived metrics validate their relevance in understanding social frailty. In addition, Fitbit-derived RAR metrics were found to be associated with cognitive function, particularly memory decline, in partial alignment with the proposed hypothesis.

The results of the analyses using the multinomial logistic regression models partially supported the primary hypothesis. Among the step-based rhythm metrics, only the ICV was significantly associated with social prefrailty. On the other hand, the dRHR and UpMesor showed significant associations with both social frailty and prefrailty. Furthermore, the α parameter and the SD of the heart rate were significant predictors of social prefrailty. Notably, except for the ICV, all the significant predictors of social frailty were derived from heart rate data.

First, the dRHR showed significant negative associations with both social prefrailty (OR 0.82, 95% CI 0.68-0.97; P=.02) and social frailty (OR 0.74, 95% CI 0.58-0.94; P=.02) in model II. The dRHR reflects the difference between the daily average heart rate and the RHR, which may provide insights into the activity levels and physiological responses to rest and activity. In this study, we found that the dRHR values were smaller in the social frailty and prefrailty groups, potentially indicating reduced daytime activity levels and a closer approximation of the average heart rate to the RHR.

Figure 2 illustrates our finding of lower daytime heart rates and higher nocturnal heart rates in the social frailty group. This pattern corresponds to smaller dRHR values, suggesting a possible disruption in the typical fluctuations between rest and activity. Notably, insufficient nocturnal heart rate reduction in persons with social frailty might contribute to an elevated RHR. In addition, increased physical activity has been associated with reduced RHRs, likely mediated by autonomic nervous system regulation [34]. Therefore, lower dRHR values might reflect both reduced daytime activity and limited heart rate reduction during rest. In this context, dRHR can be interpreted as a metric associated with variations in heart rate across periods of activity and rest.

Studies on the association between the severity of depression and heart rate variation have suggested that reduced activity levels can lead to altered heart rate fluctuation patterns [35]. Specifically, a higher nocturnal resting mean heart rate has been associated with more severe depressive symptoms (β=0.09; P<.001), whereas greater daytime heart rate variation (β=−0.34; P<.001) and a higher SD of daytime heart rate (β=−0.18; P<.001) have been linked to a lower severity of depression [35].

Similarly, previous research has identified a high RHR as a risk factor for dementia [36], worsening of depression [35], noncardiac disability [37], and physical frailty [38]. Social frailty has also been reported to predict disability [24], cognitive decline [39], depression [40], and physical frailty [6,39,41]. On the basis of these findings, dRHR may function as a comprehensive indicator reflecting these interconnected factors.

However, research on the dRHR is limited. Only the studies by Rykov et al [16,17] have examined dRHR in patients with heart disease and depression. Our study is the first to report on dRHR in the context of social frailty and has highlighted its relevance and potential value for understanding social frailty.

In addition, another heart rate indicator, the SD of the heart rate, was identified as a predictor of social prefrailty (OR 0.75, 95% CI 0.59-0.94; P=.02; model IV). Similarly to dRHR, the SD of the heart rate can be interpreted as an indicator of the heart rate fluctuations associated with daily physical activity levels. It is also classified as a measure of the HRV, a widely recognized method for assessing autonomic nervous system function. The SD of the heart rate has been classified as a time-domain HRV metric in previous research [42]. While the SD of the heart rate has been less frequently studied compared with other HRV metrics such as the RMSSD and SD of the NN interval, its potential relevance in the context of social frailty is noteworthy. In this study, we also calculated the RMSSD, but no significant differences were observed between groups. While HRV is a well-established metric for assessing autonomic nervous system function, its specific significance in social frailty remains underexplored. Future research should incorporate not only time-domain measures such as RMSSD, which primarily reflect parasympathetic nervous system activity, but also frequency-domain metrics, including low- and high-frequency components and their ratio (LF/HF), which reflect the balance between sympathetic and parasympathetic activity [43]. This may lead to a deeper understanding of the role of HRV in the context of social frailty.

Second, this study found that the UpMesor was delayed in the social frailty group (OR 1.45, 95% CI 1.09-1.94; P=.01; model II). Although previous studies have not used heart rate–based UpMesor, our results align with reports from previous studies that delayed activity onset is associated with depressive symptoms and functional decline [44]. Given the association between social frailty and depression [45-47], delayed activity rhythms may represent a characteristic feature of social frailty.

Conversely, the UpMesor was advanced in the social prefrailty group (OR 0.57, 95% CI 0.46-0.71; P<.001; model II). Typically, social frailty progresses linearly from robust to prefrailty and then to frailty, suggesting that UpMesor would also be progressively delayed across these stages. However, our findings present an apparent contradiction.

According to the report by Ji et al [48], UpMesor, DownMesor, and acrophase are parameters reflecting temporal rhythms, including the onset and cessation of nighttime sleep and peak daytime activity levels. The observed opposing patterns of UpMesor suggest the importance of understanding how specific temporal rhythm elements impact health across the stages of social frailty.

Previous research has reported on acrophase, revealing complex associations. For example, delayed acrophase has been linked to developing dementia or mild cognitive impairment [49], whereas early acrophase has been associated with cognitive decline [50,51]. These studies indicate that temporal rhythms do not exhibit simple linear relationships with health; both excessively delayed and advanced rhythms may negatively impact health.

The observed delay in UpMesor in the social frailty group and its advancement in the social prefrailty group may support this nonlinear association. Specifically, early activity onset in the prefrailty stage and delayed activity onset in the frailty stage may each contribute to health risks. This finding underscores the need for further research to examine how factors such as sex, age, and physiological characteristics might influence these associations.

Third, the α parameter × 100 was significantly associated with social prefrailty (OR 1.05, 95% CI 1.01-1.09; P=.02; model IV), suggesting that an increase in α may heighten the risk of social prefrailty. High α values reflect shorter activity periods and longer rest periods, corresponding to insufficient daytime activity. This aligns with the findings of Smagula et al [44], which linked high α values to increased health risks such as depression.

However, α may also indicate rhythm clarity, as suggested by Ji et al [48] using the sleep regularity, satisfaction, alertness, timing, efficiency, and duration model [48]. Therefore, α must be interpreted cautiously as it may reflect both the quantity and quality of daytime activity and nighttime rest.

Fourth, the ICV × 100 was significantly associated with social frailty (OR 1.04, 95% CI 1.01-1.08; P=.01; model IV). ICV reflects the intrahourly CV, with higher values indicating greater variability in activity levels. In this study, individuals with social frailty exhibited more frequent alternations between short activity bursts and rest, consistent with the report that reduced physical activity and limited social engagement are associated with the risk of frailty [6]. These results suggest that intradaily rhythm instability may be a key feature of social frailty.

In addition, we found a significant association between ICV and WLM score (β=−0.04; P=.02), suggesting that higher rhythm variability may be linked to lower cognitive performance. This finding aligns with those of previous research indicating that increased IV and decreased IS are associated with cognitive decline [52,53]. These results may imply that the ICV metric, which reflects short-term rhythm fragmentation, is related to cognitive function. However, as Fitbit does not record activity counts, relying solely on step-based data may lead to underestimation of activity levels in individuals with impaired mobility.

One of the significant strengths of this study is its use of Fitbit data to elucidate new rhythm-related characteristics associated with social frailty. Fitbit devices have demonstrated acceptable precision comparable to that of research-grade devices and are also affordable and accessible, making them promising tools for the development of digital biomarkers. In this study, step count and heart rate data recorded by Fitbit were used to replicate rhythm metrics based on the scripts by Rykov et al [17], and new rhythm metrics were also calculated. We demonstrated in this study that this comprehensive approach offers the potential for evaluating rhythm-related characteristics in social frailty.

Furthermore, the combined use of nonparametric metrics and extended cosinor metrics has shown that multifaceted evaluation of rhythm characteristics is achievable. Nonparametric metrics are well suited for capturing the daily regularity of rhythms, whereas extended cosinor metrics are useful for analyzing phase changes along the time axis. This complementary use uncovered new aspects of social frailty that were not previously identifiable through conventional methods.

In addition, extended cosinor metrics, which have traditionally been calculated based on activity counts, were applied to heart rate data in this study, revealing new possibilities for rhythm analysis. This development suggests that rhythm evaluations can be conducted even for individuals with reduced walking ability or limited daytime movement. The use of heart rate–based data expands the scope of analysis to groups that have been challenging to evaluate previously, providing a more comprehensive understanding of the rhythm characteristics.

Despite its strengths, this study had several limitations. First, the sample size posed a constraint. Due to the limited sample size, the findings should be interpreted as observational results. For example, determining the “optimal range” for UpMesor requires validation in larger sample sizes. Furthermore, the small sample size restricted our ability to conduct comprehensive analyses integrating nonlinear models or multiple health outcomes.

Second, sex differences may have influenced the results. The potential impact of sex on the outcomes necessitates cautious interpretation of the regression model estimates. Future research should incorporate sex considerations to provide more nuanced insights.

Third, there is the issue of uncertainty surrounding the standardization of social frailty indexes. According to a meta-analysis conducted by Jia et al [54], lack of a standardized social frailty index and reliance on operational choices by individual researchers remain significant challenges. This calls for further discussion and validation of the reliability of social frailty indexes.

Fourth, although a significant association between rhythm metrics and cognitive function (WLM score) was observed, this analysis was exploratory and secondary in nature. As the primary objective of this study was to model social frailty, these findings should be interpreted as preliminary. Advanced modeling techniques such as machine learning were not used in this study due to the limited sample size and the descriptive aim of the cognitive analysis. Future studies with larger samples should incorporate machine learning to develop predictive models for cognitive outcomes and validate the utility of rhythm-based digital biomarkers.

In addition, the interpretation of extended cosinor metrics presents a challenge. The current body of research on extended cosinor metrics remains limited, and careful consideration tailored to specific objectives and populations is required. Unlike nonparametric metrics, cosinor metrics do not allow for simple binary evaluations such as “stable” or “unstable.” Both excessive stability and instability of rhythms may be associated with health risks, which necessitates a comprehensive approach to rhythm evaluation.

Finally, there are limitations related to Fitbit data. While Fitbit and similar smartwatches provide minute-by-minute step data through their application programming interfaces, they do not provide activity count data. This lack of activity count data raises questions about the validity of rhythm metrics derived solely from step data, highlighting the need for future research to address this issue.

This study revealed associations between novel rhythm metrics derived from step count and heart rate data recorded via Fitbit devices and social frailty. These findings suggest that digital biomarkers for evaluating social frailty and its risk factors can be derived from data provided by consumer-grade wearable devices, which are affordable and accessible.

In particular, it was shown that heart rate–based metrics such as dRHR, UpMesor, and the α parameter may play a crucial role in the assessment of social frailty and prefrailty. These metrics have the potential to complement existing activity count–based rhythm analyses by capturing new dimensions of health that were previously overlooked.

Furthermore, the findings of our study suggest that social frailty is associated with specific rhythm characteristics such as decreased daytime activity levels and increased nighttime heart rates. These findings indicate that rhythm metrics may be valuable for early detection and assessment of social frailty.

This research provides preliminary insights into the evaluation of social frailty using Fitbit data. However, limitations such as the small sample size and uncertainty surrounding standardized indexes for social frailty must be addressed. Future studies should focus on validating these findings in larger samples and conducting comparative research using other devices. In addition, further exploration is needed to determine whether heart rate–based metrics can be applied to assess other health conditions and diseases.