Population differences in wearable device wear time: Rescuing data to address biases and advance health equity
1Department of Genetics, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States
2Department of Neurology, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States
3Department of Health Sciences, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States
4Department of Biomedical Informatics, University of Pittsburgh, Pittsburgh, PA, United States
5Division of Biostatistics, University of California Berkeley School of Public Health, Berkeley, CA, United States
6Department of Biomedical Engineering, Duke University, Durham, NC, United States
*Corresponding author: Eric Hurwitz, 120 Mason Farm Road, Chapel Hill, NC 27599 eric_hurwitz@med.unc.eduAbstract
Wearable devices present transformative opportunities for personalized healthcare through continuous monitoring of digital biomarkers; however, individual variations in device wear time could mask or otherwise impact signal identification. Despite the widespread adoption of wearable devices in research, no comprehensive framework exists for understanding how wear time varies across populations or for addressing wear time-related biases in analysis. Using Fitbit data from 11,901 participants in the All of Us Research Program, we conducted the first large-scale systematic assessment of wearable device wear time across demographics, social determinants of health, lifestyle factors, mental health symptoms, and disease. Our findings revealed that wear time was higher among males and increased with age, income, and education, but decreased with depressive, anxiety, and anhedonia symptoms, with reductions more pronounced following clinical diagnoses compared to symptom-based classifications. Individuals with chronic conditions displayed differential levels of wear time compared to healthy controls. Critically, we demonstrate that the widely used ≥10-hour daily compliance threshold, while appropriate for some research contexts, can disproportionately exclude days of data from disease populations: among individuals with major depressive disorder, 74.4% of data days were excluded compared to 20.9% for controls. We propose a flexible methodological framework including standard compliance thresholds, wear time covariate adjustment, metric normalization, propensity score matching, and adaptive thresholds that can be applied individually or in combination to optimize wearable data retention across diverse research contexts. These findings establish wear time as a critical methodological consideration for wearable device research and provide guidance for advancing equitable and rigorous digital health analytics.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Funding Statement
Individual authors were supported by NHGRI RM1HG010860 (EH and MAH), 1-K12-AR085544-01 (HM), NSF CAREER #2339669 (JD) and NIH R01 DK133531 (JD).
Introduction
The rise of wearable device ownership, such as Fitbits, has led to significant advancements for personal health1. Because wearables monitor the same individual in a continuous and longitudinal manner, their use is particularly exciting in the movement towards personalized medicine2. Wearable devices collect digital biomarkers to capture data passively and objectively in non-clinical settings. Digital biomarkers from wearables present an opportunity to transform healthcare and research capabilities with low-cost and convenient technology to provide personalized insights about health3.
Prior large-scale longitudinal, observational studies have shown physical activity4 and sleep patterns5 are associated with incident chronic disease. Additionally, machine learning using digital biomarkers from wearables have clinical applications ranging from acute event detection (e.g., infection6–8) to management of chronic disease (e.g., cardiovascular disease9,10 and diabetes11,12). While these studies established a relationship between sensor-derived digital biomarkers (e.g., heart rate, physical activity, and sleep) and disease through statistical and machine learning methods, how to methodologically account for individual variations in device wear time, which could mask or otherwise impact signal identification, remains unresolved. Prior studies looking at wearable device wear time have mainly taken place in general, disease-agnostic populations, and focus on non-biomedical contexts to study user behavior with technology13–18. It is crucial to investigate wearable device wear time across participants with a spectrum of health and disease states to maximize the clinical utility of wearable devices.
Applications of wearable devices and their data are clinical trials, observational studies, and population health research, where understanding wear time patterns is essential for maximizing the return on data collection investments. To maximize the utility of these data, it is necessary to consider not only sensor-derived biomarkers (e.g., heart rate, physical activity, and sleep) but also wear time itself, which may serve as an indicator of clinical status and outcomes19. Currently, digital health studies typically filter on ‘compliant’ data days (≥10 hours wear time), which discards data that can still contribute valid signal when accounted for and creates sampling bias against groups with lower device adherence, while also incurring financial costs from data exclusion. The ≥10-hour threshold originates from the NHANES accelerometry literature, where it was established to ensure sufficient data capture for reliable estimation of daily activity patterns in population surveillance studies. While this criterion remains appropriate for certain applications, particularly when capturing a representative single-day snapshot of behavior in healthy populations, its universal application to disease-specific research may inadvertently discard clinically meaningful data18. In clinical contexts, reduced wear time itself may reflect disease burden, symptom severity, or treatment effects, making these observations particularly valuable rather than expendable. For example, a study including wearable devices in breast cancer patients undergoing chemotherapy suggested that drug toxicity may contribute to decreased engagement with the device20. Additionally, an observational study using wearable devices, the electronic Framingham Heart Study, found that individuals with depressive symptoms were likely to have lower wear time of their device21. Overall, levels of wearable device wear time is likely non-random and can offer valuable insights in the context of study participants during data collection.
This study aimed to systematically characterize how wearable device wear time varies by demographics, social determinants of health (SDoH), lifestyle factors, mental health symptoms, and a range of chronic disease categories to establish wear time as a critical methodological consideration for wearable device research. We propose a flexible analytical framework for maximizing wearable device data retention, including adjusting for hours of daily wear time as a covariate alongside complementary approaches such as metric normalization, propensity score matching, and adaptive wear time thresholds. We leveraged Fitbit, survey, and electronic health record (EHR) data from the All of Us (AoU) research program, an initiative that aims to gather multiple health-related streams of information in one million Americans with a focus on patient populations that are under-represented in research22. This comprehensive approach provides both a foundational characterization of wear time patterns across diverse populations and practical methodological guidance for optimizing wearable device research while advancing health equity.
Methods
Data source and platform
Data described in this study were analyzed using the AoU Controlled Tier v7 dataset. All analyses were conducted in the Researcher Workbench cloud platform using R. Fitbit data in AoU v7 is collected under a bring-your-own-device model, where study participants who already own a Fitbit share their data by linking their device to the AoU portal.
Ethical considerations
The protocol for the AoU study was reviewed by the institutional review board of AoU (protocol 2021-02-TN-001). The institutional review board follows the regulations and guidance of the National Institutes of Health Office for Human Research Protections for all studies, ensuring that the rights and welfare of research participants are overseen and protected uniformly. The informed consent process states that participants have the option to withdraw at any time. Privacy of participant data is maintained in the following three ways: 1) storing data on protected computers, 2) preventing researchers from seeing identifiable patient information, such as name or social security number, 3) having researchers sign a contract they won’t try to identify participants. Furthermore, access to the AoU dataset is only available through the Researcher Workbench, which is only available to researchers who have completed the requisite training at institutions with a signed Data Use Agreement. For compensation, participants are offered $25 one-time in the form of cash, gift card, or an electronic voucher if they are asked and decide to go to an AoU partner center for physical measurements to give blood, saliva, or urine samples.
Statistical analyses
We assessed Fitbit wear time using two key research questions: 1) “How does the percentage of days individuals wear their Fitbit differ between groups X and Y?” and 2) “What is the difference in the daily hours of Fitbit wear time between groups X and Y?”. Analysis was performed in the same manner as our prior work23 and is described below.
For question 1, we calculated the percentage of days each individual wore their Fitbit using the number of days an individual had any recorded step data divided by the total number of days they could have worn their Fitbit (i.e., the difference in days between the first and last day each person had any record of Fitbit data). Analysis of Fitbit wear time was conducted using either linear regression or linear mixed-effects models using the lmer package in R24. Linear mixed-effects models were used in the case where there was more than one measurement per person to avoid the issue of correlation between data points. Please refer to Supplemental Methods in the Data preprocessing for covariates section for details about how covariates were obtained in AoU. For question 2, we determined the daily Fitbit wear time (in hours) per individual by counting the number of hours in a day with non-zero step counts, consistent with prior work4,25. To optimize computational efficiency in the Researcher Workbench, we computed mean daily wear hours per participant across relevant time windows for each analysis and applied linear regression models. Consistent with our earlier approach, we used linear mixed-effects models to account for within-participant correlation when multiple observations were present. Statistical significance was assessed at α=0.05 with Bonferroni correction applied to account for multiple comparisons within each analysis (e.g., when comparing across 6 racial/ethnic groups, the corrected significance threshold was 0.05/5=0.01). All models from both questions 1 and 2 were assessed at a significance level of 0.05 with Bonferroni correction26.
Data preprocessing and statistical analysis comparing Fitbit wear time by demographics, SDoH, and lifestyle factors
Age, sex, and race/ethnicity were obtained for each individual using the Person table in AoU. For each individual with any Fitbit data available in the AoU, we calculated 1) the percentage of days and 2) the number of daily hours they wore their Fitbit as described above. For the two calculations, we filtered on individuals with at least 30 total days of Fitbit step data (since step data was used to estimate wear time) to ensure a reliable pattern of wear time6,7. For instance, if someone had only 2 days of available Fitbit data and wore their Fitbit both days, they would show 100% wear time which would be misleading due to the limited data. When preparing data for the number of daily hours each individual wore their Fitbit, we calculated the average hours of wear time per person across time periods of interest and then performed linear regression models. Race/ethnicity was determined by merging the race and ethnicity columns in the Person table. Furthermore, we recategorized individuals’ race/ethnicity groups with less than 20 participants as Other NH race/ethnicity in order to comply with AoU’s requirement of at least 20 participants per group to mitigate the risk of participant reidentification27.
Preparing and analyzing mood, anxiety, and anhedonia symptom survey data
Survey questions about individual levels of mood, anxiety, and anhedonia symptoms were assessed using all six versions deployed from May 2020 through February 2021 of the COVID-19 Participant Experience (COPE) survey28. Please refer to Supplemental Methods in the Preprocessing depressive, anxiety, and anhedonia symptoms for details about how data were obtained and prepared in AoU.
Preparing and analyzing EHR data amongst those with major depressive disorder (MDD) and anxiety disorder (AD)
Individuals were identified as part of the MDD cohort following the Depression phenotype workspace available in AoU29. Please refer to Supplemental Methods in the Preprocessing major depressive disorder and anxiety disorder diagnosis data for details about how data were obtained and prepared in AoU.
Preprocessing and analysis of individuals with mental disorders
Individuals were classified as having a mental disorder if they had any EHR diagnosis of at least one mental disorder listed in Table S1 based on the presence of an EHR diagnosis using the Conditions table in AoU. Groups of mental disorders included were selected based on the Phemap and Phecodes as described previously4,5,30. Specific mental disorders were excluded from analysis if there were less than 20 individuals to stay in compliance with the AoU dissemination policy27. For each individual, we first identified the date of diagnosis for each mental disorder and then determined the percentage of days their Fitbit was worn between the date of diagnosis for that mental disorder and the last recorded day of any Fitbit data in AoU. To ensure that the data analyzed was during when individuals had the mental disorder, Fitbit data was filtered after the index date for each mental disorder. For individuals diagnosed with mental disorders before Fitbit devices became commercially available in 2009, we reset the index date to January 1, 2010. A cohort of individuals without mental disorders was created by excluding those who did not have a diagnosis of any mental disorder listed in Table S1, where the index date for these individuals was the median number of days each disorder was diagnosed after the first day of Fitbit data in the cohort with mental disorders, consistent with our prior work25. To temporally align comparisons between individuals with and without mental disorders, we calculated the median index date among individuals with mental disorders (July 19, 2016) and assigned this date as the index date for control individuals without mental disorders. We also filtered on individuals with at least 30 total days of Fitbit data (in total and with their mental disorder) to ensure a reliable pattern of wear time. We then performed a linear mixed-effects model with person ID as the random effect (since some individuals could have more than one mental disorder) to compare the 1) percentage of days and 2) hours per day the Fitbit was worn between those without a mental disorder and individuals with each mental disorder at a significance level of 0.05 with Bonferroni adjustment and adjusted for age, sex, race/ethnicity, annual income, education level, and health insurance status.
To compare differences in the percentage of Fitbit wear time before and after the diagnosis of each mental disorder, we determined the possible number of days each individual could have worn their Fitbit between the first recorded day of any Fitbit data in AoU and the date of mental disorder diagnosis, counted the number of days each individual had any step data between those two dates, and calculated the percentage of days each individual wore their Fitbit before diagnosis. This process was repeated in a similar manner to determine the percentage of days each individual wore their Fitbit after being diagnosed with a mental disorder, anchoring between the date of diagnosis and the last date of any recorded Fitbit data for that person in AoU. A similar process was conducted to determine the number of hours per day. The same methods were performed for each mental disorder. Linear mixed-effects models with person ID as a random effect tested the interaction between mental disorder diagnosis and time period (before/after index date) on Fitbit wear time 1) percentage and 2) hours per day, adjusting for age, sex, race/ethnicity, annual income, education level, and health insurance status with Bonferroni correction.
Preprocessing and analysis of individuals with different types of disorders
Individuals were identified as having one or multiple of the chronic disorders listed in Table S2 based on the presence of an EHR diagnosis using the conditions table in AoU. All groups of disorders included were selected based on the Phemap and Phecodes4,5,30. Disorders with counts of less than 20 individuals were excluded due to AoU privacy policies27. We also excluded Congenital anomalies because these are present from birth, making it impossible to compare Fitbit wear time before versus after diagnosis. We followed the same steps as described with mental disorders to determine the percentage of days and hours per day each individual wore their Fitbit with each disorder and compared that to a healthy control cohort, defined as someone who was not diagnosed with any of the individuals without any of the conditions listed in Table S2. The analysis was done using a linear mixed-effects model with person ID as the random effect adjusted for age, sex, race/ethnicity, annual income, education level, and health insurance status at a significance level of 0.05. We also assessed differences in wear time before and after the diagnosis of each condition using an interaction model as described above with Bonferroni correction.
Wear time adjustment methods
Variability in device wear time across individuals and clinical populations presents a methodological challenge for wearable device research, as standard compliance filtering may systematically exclude data from populations with reduced wear adherence. To address this, we evaluated five analytical approaches for handling wear time variability, each with distinct assumptions, strengths, and limitations (Table 1). These methods are not mutually exclusive and can be selected or combined based on the research question, outcome of interest, and study population. Previous evaluations of observational health data have addressed similar concerns regarding heterogeneous monitoring31–34.
To demonstrate how each approach affects estimation of a well-established clinical relationship, we compared daily Fitbit step counts between individuals with and without MDD, a condition selected because reduced physical activity is a well-documented feature of depression. All models were fit as linear mixed-effects models with a random intercept for person ID and adjusted for age, race/ethnicity, and sex at birth. The five approaches were implemented as follows. First, standard compliance filtering retained only person-days meeting established quality criteria (≥10 hours of wear time, >100 steps, and <45,000 steps) and modeled total daily steps as the outcome. Second, covariate adjustment retained all person-days regardless of wear duration and included hours of daily wear time as an additional covariate, with total daily steps as the outcome. Third, metric normalization retained all person-days with non-zero wear time and used steps per hour of wear time as the outcome, directly standardizing activity by wear duration rather than adjusting for it statistically. Fourth, propensity score matching used nearest-neighbor 1:1 matching without replacement on hours of wear time at the person-day level to create comparison groups with equivalent wear time distributions, then modeled total daily steps on the matched sample. Fifth, adaptive threshold analysis systematically evaluated minimum wear time thresholds from 1 to 10 hours, at each threshold filtering the data accordingly and testing whether wear time remained significantly different between groups using Wilcoxon rank-sum tests, providing a sensitivity analysis of how threshold selection affects both data retention and residual confounding by wear time. Rather than selecting a single optimal threshold, this approach illustrates how researchers could calibrate a study-specific threshold at which wear time no longer differs significantly between comparison groups, though in the present MDD example, wear time differences persisted across all thresholds tested.
Creating graphs
All graphs were created using the ggplot2 package35 in R.
Large Language models (LLMs)
ChatGPT (GPT-4o and GPT-3.5), developed by OpenAI36, and Claude (Sonnet 4.5, Opus 4.5), developed by Anthropic37, were used to edit some portions of the manuscript, including grammar, language, and synonyms. All recommendations from LLMs were reviewed by the author and were not used for the purpose of generating ideas or content.
Results
AoU cohort characteristics
Our AoU analysis included 11,901 individuals with available Fitbit and survey data (Figure 1). Of those, the majority of individuals were female (n=8,368, 70.3%). By age and race/ethnicity, the largest subgroups in our cohort were between 60-70 years old (n=2,666, 22.4%) and White non-Hispanic (NH) (n=9,665, 81.2%; Table 2). When examining SDoH, the majority of those in our cohort had an annual income between $100-150K (n=2,485, 20.9%), held an advanced degree as their highest level of education (n=4,634, 38.9%), and had health insurance coverage (n=11,602, 97.5%; Table 2). In contrast, the smallest groups included those who skipped the annual income survey question (n=224, 1.9%), skipped the highest education level survey question (n=160, 1.3%), and those without health insurance (n=148, 1.2%; Table 2). In regards to lifestyle, the majority of individuals have not smoked more than 100 cigarettes (n=7,369, 65.8%) and have consumed at least one alcoholic beverage (n=10,694, 95.4%) in their lifetime (Table 2).
Wearable device wear time varies by demographics and social determinants of health
We first aimed to assess differences in Fitbit device wear time across the entire AoU cohort by demographics (sex, age, race/ethnicity), SDoH (income, education, health insurance status), and lifestyle factors (cigarette and alcohol use). This analysis was conducted by comparing the percentage of days the Fitbit was worn across different groups. The results showed a significant difference in the percentage of days Fitbits were worn with increasing wearable wear time with increasing age and the highest percentage of wear time among individuals aged 60-70 years old (Figure 2A). Furthermore, males exhibited increased wear time relative to females (estimate=2.83%, 95% confidence interval [CI]=1.82, 3.84%, P<0.001) while those who were considered other sex at birth did not exhibit different wear time relative to females (estimate=0.17%, 95% CI=-4.03, 4.36%, P>0.05) (Figure 2B). Individuals who were Hispanic/Latino of any race displayed decreased wear time compared to White NH counterparts (estimate=-6.19%, 95% CI=-8.10, –4.28%, P<0.001; Figure 2C). Regarding income and education, individuals with higher income and higher education had increased percentages of wear time (Figures 2D and 2E). We also detected that those with health insurance wear their device significantly more than those without health insurance (estimate=6.54%, 95% CI=2.42, 10.7%, P<0.01; Figure 2F). When assessing lifestyle differences, the percentage of wear time did not differ significantly between individuals who smoked or consumed alcohol and those who did not. (Figures S1A and S1B).
We then aimed to compare variations in wearable device wear time among these groups, measured by the number of hours per day the device was worn. We found 1) individuals who were 20-30 wore their device fewer hours than those who were 60-70 (estimate=-0.68 hours, 95% CI=-0.89, – 0.47, P<0.001), 2) males wore their Fitbit more hours per day than females (estimate=0.68 hours, 95% CI=0.55, 0.81, P<0.001), and 3) individuals who were Black NH and Hispanic/Latino of any race wore their Fitbit fewer hours per day than those who were White NH (Black NH: estimate=-0.47 hours, 95% CI=-0.74, –0.20, P<0.01; Hispanic/Latino of any race: estimate=-0.49 hours, 95% CI=-0.73, –0.24, P<0.001; Figures 3A-3C). With regards to SDoH features, we showed 1) individuals with a higher income wore their device more hours per day, 2) individuals with an advanced degree wore their Fitbit more than those whose highest completed education was grade 12, GED, or lower, and 3) there was no difference in the number of hours per day individuals with and without health insurance wore their Fitbit (Figures 3D-3F). When evaluating hours per day of Fitbit wear time by smoking and alcohol use, we found no significant difference between individuals who smoked or drank compared to those who did not (Figures S2A and S2B).
Depressive, anhedonia, and anxiety symptoms are associated with reduced Fitbit wear time
Previous studies have suggested that mood and anxiety symptoms can influence factors related to user engagement, such as the amount of time individuals wear their wearable devices21. Therefore, it was hypothesized that those with depressive and/or anxiety symptoms would wear their device fewer days and fewer hours than those without. To probe this hypothesis, we used three questions from the survey data in AoU asking about one’s level of 1) depressive, 2) anxiety, and 3) anhedonia (the inability to feel pleasure or enjoyment from activities that are typically pleasurable) symptoms over the previous two week time period. We found that individuals with depressive symptoms wore their Fitbit a lower percentage of days and no difference in the hours per day than those without depressive symptoms (percentage: estimate=-6.00, 95% CI=-8.20, –3.81, P<0.001; hours: estimate=-0.25, 95% CI=-0.53, 0.02, P>0.05) (Figure 4A). The same pattern persisted in individuals with and without symptoms of anxiety (percentage: estimate=-3.74, 95% CI=-6.05, –1.43, P<0.01; hours: estimate=-0.08, 95% CI=-0.36, 0.20, P>0.05) (Figure 4A). Individuals with symptoms of anhedonia wore their Fitbit a fewer percentage of days and hours than those without anhedonia (percentage: estimate=-7.47, 95% CI=-9.67, –5.28, P<0.001; hours: estimate=-0.31, 95% CI=-0.58, –0.03, P<0.05)(Figure 4A). When analyzing multinomial survey responses with more granularity, we observed a dose-dependent trend with decreased percentage of wear time among those with more severe depressive, anxiety, and anhedonia symptoms; however there was no difference in the hours of wear time between those without any symptoms and those with any symptoms (Figure 4B). Lastly, given the common co-occurrence of depressive and anxiety symptoms, we directly compared the percentage and hours of wear time among individuals with depressive symptoms alone, anxiety symptoms alone, or both symptoms together compared to those with neither symptom group. The results demonstrated that individuals with depressive symptoms wore their Fitbit fewer days than those with neither symptom group (estimate=-7.22%, 95% CI=-12.6, –1.87, P<0.05), those with both symptoms wore their device significantly fewer days than those with neither symptom group (estimate=-6.70%, 95% CI=-9.35, –4.04, P<0.001), and there was no difference in the hours of wear time between those with depressive, anxiety, or both symptoms compared to those without any symptoms (Figure 4C).
Individuals with depressive and anxiety diagnoses wear their fitbit less than those with neither depressive nor anxiety diagnoses
To better elucidate the association between depression and anxiety with Fitbit wear time, we leveraged the AoU EHR data to generate a cohort of individuals diagnosed with major depressive disorder (MDD) and/or anxiety disorder (AD). First, we compared the percentage and hours of daily Fitbit wear time amongst those with and without a diagnosis of MDD and found that those with MDD wore their Fitbit less than those without MDD (percentage: estimate=-59.7, 95% CI=-61.0, –58.4, P<0.001; hours: estimate=-9.59, 95% CI=-9.85, –9.33, P<0.001; Figure 5A). Similar findings were observed among those with AD (percentage: estimate=-56.5, 95% CI=-57.7, –55.2, P<0.001; hours: estimate=-9.03, 95% CI=-9.27, –8.78, P<0.001; Figure 5A). Consistent with our methods with depressive and anxiety symptoms in the survey data, we combined and assessed the relationship between Fitbit wear time among individuals with 1) both MDD and AD, 2) MDD only, 3) AD only, and 4) neither. The findings showed that individuals in groups 1, 2, and 3 wore their Fitbit significantly fewer days and hours per day than those with neither MDD nor AD (group 4; Figure 5B).
Mental disorder diagnoses show disorder-specific associations with Fitbit wear time
To expand upon our results of wearable wear time amongst those with MDD or AD, we wondered whether decreased Fitbit wear time would persist among individuals with other types of mental disorder diagnoses. Mental disorders were categorized using phecodes (Table S1), with 3,335 (28.0%) individuals having at least one mental disorder. We first examined whether there was a difference in the percentage of days individuals with each mental disorder wore their Fitbit compared to those without any mental disorder. Our results showed individuals with mood disorders, anxiety disorders, and tobacco use disorder wore their Fitbit a fewer percentage of days and hours than those without any mental disorder, while those with neurological disorders wore their Fitbit a higher percentage of days and hours compared to those without any mental disorder (Figure 6).
To examine Fitbit wear time more closely with the incidence of each mental disorder, we aimed to determine whether the percentage of Fitbit wear time differed within individuals after they were diagnosed with a mental disorder compared to before their diagnosis. To answer this question, we calculated and compared the percentage of days each individual wore their Fitbit before and after the date of diagnosis. The same analysis was performed in the AoU cohort without any diagnosed mental disorders as a control. Among individuals without a mental disorder (controls), there was a decrease in wear time after the index date (mean estimate=-7.74, 95% CI=-14.68, –0.80, P<0.05) (Figure S3). Compared to controls, there was no significant difference in the percentage of days or hours per day worn after the index date among individuals with any type of mental disorder (Figure S3). While we did observe individuals with conduct disorders exhibited a larger decline in wear time percentage after diagnosis compared to controls, this was not significant after Bonferroni correction (interaction: –19.9%, P=0.03, ns after Bonferroni correction) (Figure S3). Examining hours per day worn, individuals without mental disorders showed no significant change after the index date (estimate=-0.76 hours, P>0.05) (Figure S3). Additionally, no diagnostic groups differed significantly from controls in baseline wear time (all main effects P>0.05) (Figure S3).
Chronic conditions show disease-specific associations with Fitbit wear time
We examined whether there existed different patterns of Fitbit wear time by classes of disease compared to healthy controls (i.e., individuals without any Phecode-grouped recorded disorders). Disease classes were categorized using phecodes (Table S2), with 6,161 (52.0%) individuals having at least one chronic disease. We found that those with blood disorders, dermatologic disorders, infectious diseases, injuries and poisoning, mental disorders, musculoskeletal disorders, neoplasms, conditions during the perinatal period, and symptoms groups wore their Fitbit significantly more than healthy controls (note the control group in this analysis is healthy controls as opposed to individuals without any mental disorder in the prior analysis) (Figure 7). When comparing hours of wear time, we found that individuals with blood disorders, infectious diseases, and conditions during the perinatal period wore their Fitbit significantly more than healthy controls (Figure 7).
Similar to our prior analysis, we compared the percentage of wear time before and after the diagnosis of each class of disorders. The control group (individuals without chronic conditions) demonstrated a 5.16% decline in wear time after the index date (95% CI=-6.31, –4.01, P<0.001), consistent with natural temporal trends in device engagement (Figure S4). After the index date, individuals with genitourinary disorders (estimate=-3.19%, 95% CI=-4.74, –1.65, P<0.01, estimate=-0.51 hours, 95% CI=-0.77, –0.25, P<0.01), injury and poisoning (estimate=-3.34%, 95% CI=-5.03, – 1.66, P<0.01, estimate=-0.64, 95% CI=-0.93, –0.36, P<0.001), mental disorders (estimate=-0.42 hours, 95% CI=-0.69, –0.14, P<0.05), musculoskeletal disorders (estimate=-2.61%, 95% CI=-4.15, –1.06, P<0.05, estimate=-0.47 hours, 95% CI=-0.73, –0.21, P<0.01), and respiratory disorders (estimate=-3.43%, 95% CI=-4.97, –1.90, P<0.001; estimate=-0.59 hours, 95% CI=-0.85, –0.33, P<0.001) all wore their device fewer hours per day than controls (Figure S4).
Wear time handling methods reveal differential bias and data retention across analytical approaches
As a practical application, we demonstrated how different analytical approaches to handling wear time affect the estimation of a well-studied clinical relationship: daily step count differences between individuals with and without major depressive disorder (MDD), which was selected because reduced physical activity is a well-documented feature of depression. Current practice in wearable device research commonly addresses wear time variability by excluding ‘subthreshold’ days that fail to meet minimum wear criteria; however, if non-compliance is non-random and associated with the condition under study, this exclusion introduces selection bias that could distort clinical effect estimates. We compared five methodological approaches: 1) the standard compliance threshold (≥10 hours of wear time, >100 steps, and <45,000 steps); 2) covariate adjustment, using all days of data while including hours of daily wear time as a covariate; 3) metric normalization, calculating steps per hour of wear time rather than total daily steps; 4) propensity score matching based on hours of wear time at the person-day level; and 5) adaptive thresholds, systematically evaluating minimum wear time thresholds from 1 to 10 hours.
Across all five approaches, individuals with MDD took fewer steps than those without MDD, though the magnitude and statistical significance of the difference varied by method (Figures 8A and 8B). The standard compliance threshold yielded the largest estimated group difference (999 steps/day, 95% CI=727, 1,271, P<0.001), while covariate adjustment produced a smaller but highly significant difference (796 steps/day, 95% CI=551, 1,041 P<0.001), suggesting that a portion of the apparent activity gap under standard filtering reflects unaccounted wear time variation rather than true differences in physical activity (Figure 8A). Metric normalization confirmed reduced activity among those with MDD on a rate-adjusted scale (55 fewer steps/hour, 95% CI=34, 76, P<0.001; Figure 8A). Interestingly, propensity score matching produced an estimate in the same direction but was not statistically significant (402 fewer steps/day, 95% CI=-25, 828, P=0.065), which may be attributed to severe data loss from day-level matching (93% of data excluded), which substantially reduced statistical power and selected an unrepresentative subset of low-wear-time control days (Figure 8A). Adaptive threshold analysis revealed that the MDD versus no MDD step difference remained significant across all thresholds from 1 to 10 hours (all P<0.001), and critically, that wear time remained significantly different between groups at every threshold, including 10-hours (P<0.001; Figure 8B). This finding demonstrates that compliance filtering alone does not control for residual confounding wear time differences between clinical populations.
Each method retained different amounts of data, with important implications for statistical power and representativeness. The standard compliance threshold retained 485,232 days (72.8% of total data), but this filtering disproportionately excluded data from the MDD cohort: 58,732 days (74.4%) of MDD data were removed compared to 122,936 days (20.9%) from the no MDD cohort (Table 3). Metric normalization retained 525,212 days (78.8%), excluding only days with zero recorded wear time (Table 3). Propensity score matching retained just 47,520 days (7.1%), with the severe data loss driven by 1:1 day-level matching against a control pool roughly 22 times larger than the MDD group, producing an unrepresentative subset that substantially reduced statistical power (Table 3). Covariate adjustment retained all 666,900 days (100%), including an additional 181,668 days that would have been discarded under the standard threshold (Table 3). Adaptive threshold analysis further demonstrated that filtering disproportionately excluded MDD data at every threshold from 1 to 10 hours (Figure 8B).
Discussion
Wearable devices present transformative opportunities for personalized healthcare through continuous monitoring of digital biomarkers; however, our findings demonstrate that individual variations in device wear time represent a critical and previously undercharacterized source of methodological bias in biomedical wearable device research. Using Fitbit data from approximately 11,000 participants in the All of Us Research Program, we conducted the first large-scale systematic assessment of wear time across demographics, social determinants of health, lifestyle factors, mental health symptoms, and disease, revealing that wear time was non-randomly distributed across populations: it was higher among males and increased with age, income, and education, but decreased with depressive, anxiety, and anhedonia symptoms, with reductions more pronounced following clinical diagnoses compared to symptom-based classifications. Individuals with chronic conditions displayed differential levels of wear time compared to healthy controls. Critically, we demonstrate that the widely used ≥10-hour daily compliance threshold can disproportionately exclude data from disease populations, with 74.4% of data days excluded among individuals with major depressive disorder compared to 20.9% for controls. To address this, we propose and empirically evaluated a flexible methodological framework, including covariate adjustment, metric normalization, propensity score matching, and adaptive thresholds, that can be applied individually or in combination to optimize data retention while maintaining analytical rigor. These findings establish wear time as both a methodological consideration that must be addressed in wearable device research design and a potential informative signal in its own right.
Several factors are known to impact wearable device wear time, including age, education, income, occupation, fitness levels, and health conditions38. We observed that males, older adults, and those with higher income and education status tended to wear their Fitbit more than their counterparts. Although younger adults demonstrate higher digital health literacy and wearable adoption rates39, among individuals who have already adopted a device, older adults wore their Fitbit more consistently, consistent with previous studies40. Previous research suggests that wearable devices support “active aging”, defined as enhancing quality of life throughout the aging process41. Qualitative analysis indicated older adults are motivated to learn about their health, highlighting a key value of wearables. Research also suggests that older adults with specific needs, such as fall risk, are more likely to engage with their devices as a result of that motivation42,43. We were also surprised to find males displayed higher wear time than females despite the AoU cohort being predominantly female (70%). This may reflect self-selection bias, where males who chose to enroll in a research study and share their Fitbit data represent a more health-motivated and physically active subgroup than the general male population, a pattern consistent with research showing males who engage with wearable device features tend to be more physically active than their female counterparts44. While the difference was statistically significant, the absolute magnitude was very small. Notably, this finding contrasts with self-reported data suggesting women are more likely to use wearable health devices than men45, further supporting the idea that the bring-your-own-device enrollment model may have introduced motivational self-selection that disproportionately filtered the male sample46,47.We also observed individuals with higher incomes and education levels engage more with wearable devices, consistent with research demonstrating that children from families with higher education and income show greater device adherence48. This may be due to systemic factors, where individuals of lower socioeconomic status prioritize more immediate needs over wearable technology, limiting their device wear time and preventing them from fully experiencing its benefits. Biases in wearable device deployment and data collection exist3,49–51 and structural and systemic barriers, such as concerns about data privacy, accuracy, language barriers, and access to health care technology, impact use of wearables among racial and ethnic minorities46,47,52–55. Recognizing these disparities underscores the importance of both equitable device deployment and analytical approaches that account for wear time variability, ensuring data from lower socioeconomic populations is preserved rather than systematically excluded through strict compliance criteria47.
Our findings also showed individuals with depressive, anxiety, and anhedonia symptoms exhibited lower levels of wearable device wear time compared to those without symptoms. This is consistent with prior work suggesting depressive and anxiety symptoms are linked to lower healthy lifestyle scores, which may contribute to decreased motivation for health-tracking behaviors21,56. Lower levels of wear time were more pronounced in individuals with diagnosed depression and anxiety disorders compared to those reporting symptoms without formal diagnoses, likely reflecting greater symptom severity in the diagnosed group. Wear time was consistently reduced in those with any diagnosed mental disorder compared to those without, with notable variation across specific disorders: individuals with tobacco use disorder showed lower wear time, consistent with broader patterns of reduced health-promoting behaviors and sedentary lifestyle in this population57, while those with neurological disorders showed higher wear time, potentially reflecting the particular clinical utility of continuous device monitoring in this group58. Expanding this analysis to chronic disorders more broadly revealed similarly variable patterns; while detailed interpretation of individual conditions is beyond the scope of this paper, future research should explore specific disorders in greater depth, considering factors such as disease severity and condition-specific characteristics to better contextualize wear time patterns.
When examining differences in wear time across individuals with and without MDD in our example, we observed that those with MDD were disproportionately impacted by filtering on compliance, with nearly 75% of days of data being excluded, compared to those without MDD who only had about 21% of days excluded (which is still sizeable). This disproportionate data loss has critical implications for statistical analyses and machine learning models. By incorporating hours of wear time as a covariate rather than applying strict compliance filters, we can rescue these otherwise excluded days of data. This approach not only increases statistical power, particularly important for chronic conditions disproportionately affected by wear time, but also reduces selection bias by retaining data from vulnerable populations who would otherwise be systematically excluded from analyses. Importantly, the two methods produce different absolute activity estimates where compliant-only data yields higher step counts because low-activity days are disproportionately excluded, a consideration for downstream clinical and public health applications where accurate activity estimates are needed. Our results demonstrate that this methodological shift enables more inclusive and representative digital health research while maintaining analytical rigor.
We acknowledge that the ≥10-hour wear time threshold has sound methodological origins and remains useful for specific applications, such as estimating representative daily activity in healthy populations18. However, for disease-specific research, we propose that reduced wear time is often informative rather than simply indicative of invalid data. Several methodological approaches can address variability in wear time while preserving these observations. When the primary goal is maximizing statistical power and sample representativeness, including hours of wear time as a covariate retains all available data while statistically controlling for differential wear across groups, making it particularly suited for group-level comparisons of activity outcomes where data loss would disproportionately affect one group59. An additional consideration is the timing of available data, as systematic missingness during specific hours can attenuate or exaggerate circadian patterns. When the outcome of interest is an activity metric that scales directly with wear duration, such as step counts, normalizing by wear time (e.g., steps per hour) provides an intuitive rate-based measure that is inherently comparable across participants with variable wear, though this approach assumes a relatively constant activity rate throughout the day and requires a minimum wear time floor to avoid unstable estimates59. When the analytical design requires equivalent device exposure between comparison groups, such as in case-control studies, propensity score matching on wear time ensures balanced distributions prior to analysis, though this comes at the cost of reduced sample size and generalizability, as demonstrated by the substantial data loss observed in our MDD example60. When researchers seek to retain more data than the standard threshold while still ensuring a minimum level of data quality per day, an adaptive threshold calibrated to the study population can balance data retention with analytical validity, particularly when wear time differences between groups diminish at lower thresholds61. These approaches are not mutually exclusive; for example, a researcher could apply an adaptive threshold to ensure minimum data quality and then include residual wear time variation as a covariate. The optimal combination will depend on the research question, outcome of interest, and study population, and we encourage researchers to report results under multiple approaches as a form of sensitivity analysis. Future studies comparing these approaches against gold-standard activity measures (e.g., direct observation) would help establish validity benchmarks for different research contexts. By offering a flexible methodological framework rather than a single prescriptive solution, we aim to equip researchers with practical tools for maximizing wearable device data retention while maintaining analytical rigor across diverse clinical populations (Table 1).
Although this study is one of the first to characterize large-scale, biomedical wearable device wear time, it is not without limitations. For example, EHR-based analyses may not accurately capture disease incidence depending on where and when data are entered, as symptoms often precede both diagnosis and the subsequent recording in the EHR62,63. For this reason, we excluded certain condition groups where the first recorded date was known to be inaccurate or inappropriate (e.g., congenital disorders). Nonetheless, EHR data are still commonly used to estimate incidence, and our results should be interpreted with this limitation in mind62,63. Second, the analysis performed in this AoU cohort (v7) used a bring-your-own-device model, meaning individuals who already owned their Fitbit consented to share their data. This results in a demographic imbalance of individuals who own wearable devices, which tends to favor those of higher socioeconomic status47. Since the time of analysis, AoU updated their data to v8, which contains individuals from the WEAR study, which gave individuals from underrepresented communities a Fitbit and resulted in a new total sample size of approximately 60,000 individuals with Fitbit data in AoU64. It would be valuable for future work to replicate this study in a more heterogeneous cohort. Further, there was variability in the sample size in specific disease groups, which may limit the generalizability of our results. However, this is a common limitation of retrospective observational studies that are not collecting data for any particular population and allows us to conduct analyses across several diseases. Third, AoU has since added new Fitbit metadata, including device type and approximate battery level, which may influence wear time patterns and warrant investigation in future analyses.65. Finally, the calculation used to estimate hours of wear time, which was based on published literature4,25, relies on counting the number of non-zero steps, which may underestimate wear time by excluding periods when the individual was sedentary, although it is unusual for step counts to be exactly zero in sedentary individuals due to small movements. Future analyses could consider distinguishing between missing data (NAs) and minutes with recorded zero steps, or alternatively, use the presence of intraday heart rate measurements as an indicator of wear time40.
Overall, wear time with wearable devices varies across individuals based on a range of demographic and personal health factors. Wear time-related information can not only offer insights and serve as additional features alongside signals from wearable sensors (e.g., heart rate, physical activity, and sleep), but it can also be leveraged to maximize the amount of wearable device data available for analysis in statistical and machine learning models. Moreover, the exclusion of data based on insufficient wear time creates systematic biases, as data loss disproportionately affects specific sociodemographic and disease groups. Implementing appropriate methods and incorporating wear time data in analyses can enhance the utility and richness of longitudinal wearable device data, reinforcing the role of these devices in advancing equitable and personal health.
Supporting information
Acknowledgments
The All of Us Research Program would not be possible without the partnership of its participants. The All of Us Research Program is supported by the National Institutes of Health, Office of the Director: Regional Medical Centers: 1 OT2 OD026549; 1 OT2 OD026554; 1 OT2 OD026557; 1 OT2 OD026556; 1 OT2 OD026550; 1 OT2 OD 026552; 1 OT2 OD026553; 1 OT2 OD026548; 1 OT2 OD026551; 1 OT2 OD026555; IAA #: AOD 16037; Federally Qualified Health Centers: HHSN 263201600085U; Data and Research Center: 5 U2C OD023196; Biobank: 1 U24 OD023121; The Participant Center: U24 OD023176; Participant Technology Systems Center: 1 U24 OD023163; Communications and Engagement: 3 OT2 OD023205; 3 OT2 OD023206; and Community Partners: 1 OT2 OD025277; 3 OT2 OD025315; 1 OT2 OD025337; 1 OT2 OD025276.
Funding
Individual authors were supported by NHGRI RM1HG010860 (EH and MAH), 1-K12-AR085544-01 (HM), NSF CAREER #2339669 (JD) and NIH R01 DK133531 (JD).
Data availability
Access to the All of Us Research Program (AoURP) dataset is governed by the program’s data security and privacy policies. Eligible investigators who have completed AoURP’s required training and registered on the Researcher Workbench may request study data and code from the corresponding author.