For this study, 3 Fitbit datasets from 6 different populations were used (Table 1).

All datasets were generated by teams at Michigan Medicine at the University of Michigan, and the different data collection protocols were each approved by the University of Michigan institutional review board, under the review numbers HUM00186436, HUM00176584, HUM00184455, and HUM00185391. All participants provided informed consent prior to their participation in the study. The details of the compensation for each study are provided in Table 1.

The average wear time of the device varied by population (Figure 2). The most compliant sample was the HCT caregivers with an average daily wear time of 20.8 hours. The least compliant sample was the pediatric oncology patients with an average daily wear time of 9.8 hours. Differences were also observed in the spread of the data. For the pediatric oncology patients and caregivers and the student sample, the distribution went from 0% to above 90% of wear time. A 0% wear time indicated that participants received the Fitbit but never wore it.

There were significant differences in the average wear time between populations (Table 2). Notably, the pediatric oncology patients had significantly lower wear time than all other studied samples except for the pediatric oncology caregivers. The HCT caregivers also had significantly higher wear time than the other groups, except for the SCI and HD caregivers.

A decrease was observed in monthly wear time for all recruited groups. The average decrease was a little more than 10%, with the highest decrease recorded for the pediatric oncology patients (−19%) and the lowest decrease recorded for the HCT caregivers (−8%).

On average, across all population samples, 304 samples (minutes of walking) were needed to obtain an estimate of the walking heart rate with a standard error of 1 bpm. For most individuals, the minimal number of samples was reached in under 26 days of data collection. For the pediatric oncology caregivers, 1 participant did not have enough samples to obtain a confident estimate of the walking heart rate, but an estimate of that caregiver’s average daily step count using WearTime80 was obtained. On the other hand, 5 (17%) HCT, 3 (14%) HD, 2 (11%) SCI, 186 (10%) students, and 3 (7%) of the pediatric oncology patients did not have any valid days to calculate their average daily step count but presented enough samples to confidently obtain their walking heart rate. Figure 5 illustrates an individual’s data from the student sample within that 10%. That individual wore the Fitbit at the beginning and the end of the study for around 50% of the day. That student had 1314 samples of steps taken in a minute—enough to estimate walking heart rate—while presenting no days with a wear time higher than 80%.

The purpose of this study was to quantify how wear time could impact parameter estimates in studies using consumer-grade wearables by using real-world samples with varying participant compliance. This work is distinguished from the literature by demonstrating the conditional nature of wear time’s influence in consumer-grade wearables’ data, which varies significantly across research samples and research questions. By leveraging a broad array of datasets, including vulnerable patient populations, we showed evidence of the nuanced effect of wear time, highlighting its critical role in research using wearables, especially in samples with inconsistent wear patterns such as the pediatric oncology caregiver and patient participants used in this study. It is important to note that the pediatric oncology patients’ ages ranged from 5 to 20 years, with a mean of 11.9 years. The young age combined with illness for these subjects likely explained this population sample’s low level of wear time with large variability.

Unsurprisingly, implementing different wear time requirements to define a valid day can significantly impact parameter estimates, especially in samples with low compliance. Wear time compliance for the HD and HCT caregivers was high (Figure 6); therefore, the estimates of the samples’ average daily step counts were minimally affected by different definitions of a valid day (Figure 3). In comparison, the pediatric oncology patient sample had low wear time compliance (Figure 6); therefore, the estimate of the sample average daily step count was significantly impacted by different definitions of a valid day (Figure 3). Importantly, our analysis quantifies these differences so that other researchers can better understand the potential implications of different valid day definitions. The definition of a valid day had almost no impact on estimates of the average daily heart rate across samples. This also is not surprising as heart rate is a more stable measure that does not necessarily correlate with wear time; any biases from missing data are averaged out.

A significant novelty in this work is demonstrating the flexibility of data usage in real-world settings (ie, outside of the hospital). In fact, we showed that even if an individual’s data were not sufficient to answer a specific research question, those data could potentially still be leveraged toward a different analysis (eg, for the estimate of walking heart rate as shown in this study). In particular, the variations in the research question time frame (eg, daily, nighttime, and weekly) can lead to different levels of required compliance. For example, to accurately estimate the average number of steps per day, we need a sufficient number of days with high wear time. However, for the example we used in Aim 2—the estimation of walking heart rate—we need a sufficient number of steps taken in a bout of walking by an individual. As such, we show in our example how an individual can have a large amount of walking data throughout an entire study while not having enough individual days with high wear time (Figure 5). These findings suggest that even samples with low compliance may still be useful for answering research questions. In this case, even a day with low wear time (ie, not a valid day) may contain several valid walks useful for estimating average walking heart rate.

We also found that depending on whether one is interested in studying a population as a whole or the individuals within a population, the impact of wear time differed. For instance, in the student population, we observed only a small change in the estimated average daily step count in the population with the different methods of analysis (Figure 3). Depending on the objective of a study, that change could be negligible. However, if research is conducted to design a physical activity intervention or to evaluate the efficacy of an intervention, for example, the processing of an individual’s data without the inclusion of wear time could lead to critical errors (Figure 4). In particular, if one considers the standard classification of participants into levels of physical activity (0‐5000 daily steps: sedentary; 5000‐7500: low active; 7500‐10,000: somewhat active; 10,000+: highly active) [28], those errors can induce a misclassification. Ultimately, the intervention design or the evaluation of the intervention could be erroneous.

We quantified the phenomenon of wearables’ abandonment within our different populations (Figure 6). That phenomenon can potentially create a bias in results over time. For example, a monthly decrease in the population sample level of physical activity can be explained by an actual decline in people’s movements or by the fact that participants were not wearing their sensors as much. An appropriate quantification of wear time in the analysis can potentially remove that bias.

While the literature on accelerometer usage presents various methodologies for accounting for wear time, it is essential to acknowledge that raw acceleration data or activity counts, typical outputs from accelerometers, are not available from consumer-grade wearables. Therefore, the strategies for considering wear time in the context of these devices necessitate their own approach. Consumer-grade wearables often offer summarized or processed data, which requires tailored methods for wear time analysis to ensure accurate and meaningful interpretation in research contexts. In general, research data on the analysis of wear time and compliance with consumer-grade wearables are sparse in the literature. Often, studies using consumer-grade wearables, such as smartwatches, rely on threshold-based methodologies for processing wear time [13-16]. However, these thresholds are often implemented without sufficient justification based on the specific dataset being analyzed and the research objectives.

Claudel et al [17] developed a novel method to calculate wear time and determine a valid assessment day. Their proposed method calculated wear time using heart rate detection, step count, and the range of heart rate values. A valid day was defined as a day with more than 10 hours of wear time, with the removal of sleeping hours (from 11 PM to 5 AM). Claudel and colleagues present a structured and detailed methodology that can potentially be replicated for other studies. It highlights the importance of nighttime in wear time calculations and how researchers can conduct a sensitivity analysis to determine the impact of different nighttime limits on the outcome variables. Our study also demonstrates the potential impact of different calculations of wear time on parameter estimates. Both studies illustrate the importance of considering how the definition of a valid day may impact sample parameter estimates.

Studies have explored the factors that are related to compliance levels [13]. The NetHealth study [13] collected data from a student population using Fitbit devices and showed that some personality traits (eg, extraversion and openness) and Fitbit metrics (eg, minutes asleep and sedentary minutes) were related to compliance levels. Understanding the underlying reasons that may lead an individual to be more or less compliant can be useful in improving the overall wear time levels in these types of studies.

Another, perhaps more, direct way to increase compliance is through the use of incentives. There is a range of incentive models one can use to maintain compliance in longitudinal studies [29]. Although improving compliance allows for parameter estimates to be less sensitive to wear time, it is difficult to imagine a “one size fits all” incentive model that would work on every single participant of a study. Thus, the analysis of wear time is likely to remain a crucial step in the analysis of consumer-grade wearables data to ensure accurate parameter estimates.

In the accelerometer literature, raw outputs from the sensors are most often used to determine wear time [10,11], which removes a layer of potential error. The outputs provided by Fitbit are processed (as opposed to raw) with unknown algorithms. Here, we assumed those outputs, and in particular minute-by-minute heart rate, were reliable and could be used to accurately estimate wear time. Claudel et al [17] formulated a more complex wear time calculation algorithm fusing both heart rate detection and heart rate values revealing a potential overestimation of wear time when using heart rate detection only. Determining the most accurate estimate of wear time was beyond the scope of this study. However, further research should be conducted to validate methods for the calculation of wear time using the processed outputs of consumer-grade wearables as it may influence parameter estimates.

We assumed in Aim 2 that wear time was random. In other words, we assumed that participants did not choose a particular moment in their day to wear the device, which would bias the estimates of walking heart rate. However, this does not influence the conclusion of Aim 2, as we primarily focused on the amount of data needed to estimate a parameter with confidence, and not whether we could accurately estimate true walking heart rate.

It is important to note that the different datasets we used came from 2 different Fitbit devices (Fitbit Charge and Fitbit Inspire 2). It is possible that the versions of software and hardware are different between these 2 devices. Since the software information is proprietary information, we do not know whether the model differences impacted the used outputs. However, we expect and assume that the measurements between the 2 models are comparable and did not have a significant impact on our analysis.

Additionally, although we analyzed samples from different populations, our results may not translate to similar populations. First, each population sample was given a different set of instructions, which influenced their wear time. Second, Fitbit data syncing can be difficult and depends on the software used as well as the Fitbit model itself. Thus, the amount of data retrieved for the same population sample can be variable from one study to another. Finally, the different datasets used in this study were all collected during the COVID-19 pandemic, which may have affected the health-related behaviors of participants. We recommend that researchers conduct an analysis of wear time systematically when using sensing technologies to derive conclusions.

One of the major benefits of leveraging fitness trackers is the potential to gain a unique and objective insight into an individual’s natural behavior. However, if the analysis of wear time is not appropriately integrated into the data processing pipeline, that purported insight can become biased and imprecise, driving the results away from reality. Herein, we presented evidence of the critical impact of wear time on parameter estimates and how that impact may vary depending on the population sample, research question, study design (ie, incentives), and participant compliance. Future research could significantly benefit from our findings and methodological examples to incorporate the following recommendations: first, it is essential to conduct a thorough analysis of the wear time of the population in any given study that uses wearables, with particular attention to the data available when using consumer-grade wearables; second, it is equally important to systematically document and provide comprehensive details on that analysis. Our work provides quantitative data on the impact that different valid day definitions can have on sample parameter estimates in samples with varying wear time compliance.