A key goal of the study is to explore methods to collect a representative study sample that reflects the relative proportions of various demographic subgroups and their intersectionalities in adults of the US population. For example, according to the US Census Bureau [17], 51% of the population is female and 44% of adults are between 20 and 44 years of age. Among those who reported a single race, about 13% are African American, 6% are Asian, and 74% are White. A February 2021 Gallup poll [18] reported that 5.6% of US adults identified themselves as lesbian, gay, bisexual, or transgender; 86.7% said that they were heterosexual or straight; and 7.6% refused to answer. In the past, some subgroups have been traditionally underrepresented in health studies (eg, female individuals; African American individuals; and those with a lesbian, gay, bisexual, transgender, queer, intersex, and asexual orientation) [19], and recent reviews have emphasized the importance of setting explicit recruitment goals in overcoming these biases [19,20]. As such, a key aim of the study is to learn more about how to increase the participation of traditionally underrepresented groups in digital health studies.

To explore methods to overcome recruitment barriers of traditionally underrepresented subgroups and make sure the study sample reflects sufficient diversity, we will pursue a two-prong approach: (1) a primary-augmented 2-wave metered recruitment schema and (2) multistakeholder engagement. Metered recruitment schema refers to the potential delay of eligible study participants in consent and enrollment (baseline characterization and initiation into the monitoring portion of the study). The pace of enrollment will be adapted based on study and clinical bandwidths and the distributions of demographic characteristics, with metering instituted by invitation to consent. The research team plans to begin the metered recruitment strategy once the study has achieved or has nearly achieved the enrollment of 10,000 participants. Depending on the distribution of key demographic variables, once consent and appropriate permissions are obtained, enrollment and initiation of monitoring will commence. To help reduce the effect of seasonality and other external factors on the data collected in this study, we are careful to design the strategy to recruit participants in a balanced manner across time and outside of periods containing major seasonal holidays.

During the initial several weeks that the study is open, we anticipate enrollment of up to approximately 1000 new participants per week (but will not limit to this pace). Enrollment will increase as infrastructure allows. An increase in consecutive enrollment of up to 10,000 new participants who are eligible for study participation will be invited to begin the enrollment process. The study will monitor the progress in recruiting traditionally underrepresented groups. Recruitment efforts will focus on enriched representations, with recruitment via email, social media, and Fitbit in-app notifications. Posts will be made on social media platforms, leveraging the reach and audience of the University of Oregon and Google accounts. In-app notifications will be made in the Fitbit app, and these notifications ensure high visibility of the study. At the conclusion of wave 1, the study investigators will assess study participation rates and revise the enrollment targets for wave 2.

Upon the completion of wave 1, wave-2 efforts will discontinue the enrollment of sufficiently represented groups and focus entirely on increasing the participation of underrepresented groups, as defined by race and ethnicity (eg, African American, Asian, Latina or Latino, Native Americans or Indigenous, and White populations), biological sex at birth (female and male), age (18-40 and over 40 years), sexual orientation or gender identity (heterosexual and lesbian, gay, bisexual, transgender, queer, intersex, and asexual), and whether participants are Fitbit users. Multistakeholder engagement refers to the study’s collaboration with major stakeholders such as active groups or key opinion leaders of underrepresented groups of interest to codevelop adaptive recruitment goals and strategies.

The study will aim to recruit a study sample that reflects enriched participation of traditionally underrepresented groups in the United States. The wave-2 targeted recruitment strategy will include 5 demographic factors in addition to the ownership of Fitbit devices. If wave-1 enrollment is not sufficiently at parity with population representation, wave-2 efforts will target up to an additional 4000 participants to demographically balance the aggregate sample (potential total n≤14,000).

The study includes digital measurement from both smartphones and wearable physiological devices. After enrolling in the study, participants will be asked to grant permissions to their devices. Figure 2 shows screenshots of the phone app to illustrate the design.

Patterns of device use will be measured objectively via the study mobile app. The measurements will include:

We will collect the following data from participants who sign up with Fitbit devices. As access to devices can vary, participation is allowed with any generation of Fitbit devices. Different Fitbit devices have different sensors and algorithmic capabilities, and therefore not all the measures described below will be available for all participants. Minute-level features include accelerometer-based measures of steps and photoplethysmography-based heart rate measurements. Day-level features include accelerometer-based measures of activity (eg, total steps, number of floors climbed, and number of active zone minutes), sleep stages, duration and quality, minutely heart rate, daily resting heart rate, respiration rate during sleep, heart rate variability during sleep, skin temperature during sleep, blood oxygen saturation during sleep, and tonic electrodermal activity.

Participants will complete a battery of self-report questionnaires at the beginning (baseline) and end of the study (poststudy) as follows: (baseline only) demographics questionnaire and Big Five Inventory [21,22], Patient Health Questionnaire [21,23], Generalized Anxiety Disorder Scale [24], Patient-Reported Outcomes Measurement Information System Sleep Disturbance and Sleep-Related Impairment short form [25,26], Patient-Reported Outcomes Measurement Information System Emotional Support Short Form 4a [27,28], Shortened Smartphone Addiction Scale [29], and Perceived Stress Scale [27].

The first and last week of the study will include intensive ecological momentary assessments (EMAs). During these weeks, 3 times per day, participants will be sent 6 identical survey questions per day for 7 days, on a quasirandomized schedule. The first 5 questions will be “How X do you feel right now?” where X will be replaced with a specific affect descriptor from happy, calm, anxious, sad, stressed, and participants will be asked to rate each item. The final question is “Over the last few hours, who have you spent the most time with?” and participants will choose from alone, friends, family, spouse or partner, coworkers, and costudents. Finally, every day for the entire duration of the study, participants will be asked “In general, how have you been feeling over the last day?” once each morning.

The first phase of the hypothesis-driven data analysis will be to process the raw data signals to derive interpretable features to test explicit hypotheses. The study will test a range of specific hypotheses and it would be beyond the scope of this paper to list them all. However, we present an exemplar here to demonstrate how such hypotheses could be tested. For example, 1 hypothesis that has been proposed in the research literature is that a key mechanism by which device use may affect mental health and well-being is by disrupting healthy sleep behaviors [30]. Specifically, it has been found that device use during the prebedtime period may delay sleep onset either by displacing sleep by device use or by increasing psychological arousal prior to lights out, thereby increasing sleep onset latency. In this study, estimates of device use (in the prebedtime period) will be combined with sleep features, including bedtime and rise time, time in bed, and total sleep time. Specifically, we will test whether the hypothesized mechanisms of change (ie, sleep) mediate the relationship between device use patterns and measures of the change (slope) in measures of mental health and well-being across the study period. Mediating effects will be tested using the strategies outlined by Hayes and Preacher [31], which are standard in behavioral research trials. Using linear structural equation models, support for mediation will be determined if measures of changes in measures of mental health and well-being become nonsignificant or are significantly reduced after controlling for the hypothesized mediating variable. We will assess the joint significance of the indirect pathways (ie, the joint significance of the pathways from the predictor to the mediating variable and from the mediating variable to psychiatric outcomes) using the bias-corrected bootstrap test and will also use the bootstrapping technique to test multiple simultaneous mediators [31].

The intensive longitudinal data on these features and self-reports of well-being will be subjected to data reduction techniques such as dynamic factor analysis, which are a set of techniques for estimating common trends in multivariate time series [32]. Dynamic factor analysis is a dimension reduction technique that aims to model a multivariate observed time series in terms of a finite number of common trends, with the aim of detecting the smallest number of trends that can summarize the data without losing information. The principle is the same as in other dimension reduction techniques, such as principal component analysis and factor analysis, except that the axes are restricted to be latent smoothing functions over time. The dynamic factor analysis will initially use data from the entire 4-week data collection period to examine the patterns of association between daily patterns of device use and daily variations in well-being and then will go on to examine the data from the two 1-week burst of intensive EMA to examine within-day covariation between patterns of device use and well-being. Variables that do not load on to 1 of the identified factors, but that are of strong interest as putative risk factors based on theory or previous empirical research, will be investigated in the statistical analyses as univariate predictors.

The second phase of the statistical analyses will consist of a number of analyses to be conducted to address the aims outlined above. A range of analytic techniques will be used, including ANOVA, specialized multilevel regression techniques that have been used in previous studies of intensive longitudinal data [33], and mixed Markov models [34]. These models are appropriate for testing the hypotheses with our aims, whereby changes in well-being across time can be considered an observed state switch and analyzed using observed Markov models (also referred to as manifest or simple Markov models or Markov chains [34-36]). These methods can also be used to investigate the temporal relationship between the intensive longitudinal data streams and the optimal time lag between these measurements and state switches (ie, the time lag between a predictor and the onset of a change in well-being status), which can be investigated by using higher order models, where the current state is modeled as depending on multiple preceding states [37].

We also plan to investigate the use of neural models to create machine-learned representations in our analysis. The intensive longitudinal nature of the data collection will enable large-parameter models to be trained using unsupervised and supervised machine learning techniques. Learning with self-supervision has recently attracted increasing interest as simple pretext tasks can be used to create versatile compressed representations of data. The main advantage of these embeddings is that they provide an initialization for downstream tasks. This is a more label-efficient mechanism to train supervised models. The benefits of unsupervised or self-supervised learning techniques studying human behavior are still relatively underexplored. However, the high dimensionality and the scale of the data described in this protocol lend themselves to this type of analysis. We plan to empirically test whether leveraging these methods has a positive impact on downstream predictive performance.

Given that participants are not required to complete all surveys, missing data will be handled via established imputation or multiple imputation techniques; however, we also plan to test machine learning–based imputation using autoencoders (eg, masked autoencoders). We expect that some partial data may be unusable, with the wearable devices adherence in intensive longitudinal studies is a known issue.