Responses to mental health questionnaires were collected via monthly surveys delivered through the WHOOP mobile app to all interested WHOOP members over a 13-month period, from March 2022 to April 2023. The mental health assessment questions included the PHQ2 [16,17] (a questionnaire meant to gauge depression), the GAD2 [18] (a questionnaire meant to gauge anxiety), and the PSS [19] (a questionnaire meant to gauge perceived stress). The questions and their responses are available as Multimedia Appendix 1. To account for the different response scales between the PHQ2 or GAD2 and PSS, the PSS was linearly scaled from 0 to 6. The combined responses from the PHQ2, GAD2, and PSS assessments were then summated as a measure of overall mental state, and this score is referred to as the combination score. The scores of the PHQ2, GAD2, and PSS reflect severity or likelihood of a mental health disorder (eg, scoring higher on PHQ2 means a greater likelihood of being depressed or an increased severity of depression) [16,17,20].

A wrist-worn device (WHOOP strap version 3.0 and 4.0) that continuously collects heart rate (via photoplethysmography) and accelerometry (via 3-axis accelerometer) data was used to calculate sleep, cardiorespiratory, and physical activity. Sleep measures included total sleep duration, sleep efficiency, SD of sleep duration, sleep consistency, SD of wake times, and SD of sleep times. Sleep consistency is adapted from the sleep irregularity index and is calculated as the percentage of concordance when individuals are in the same state (sleep vs awake) at different time points over a 4-day interval, with comparisons of intervals further apart being assigned progressively lower weights [21,22]. A higher sleep consistency value reflects a lower variability in sleep-wake timing. Cardiorespiratory measures included resting heart rate (RHR), average heart rate variability (HRV_av), coefficient of variation of heart rate variability (HRV_cv), and respiratory rate. RHR is calculated as a weighted mean of heart rate during sleep. Heart rate variability (HRV) is calculated as the weighted average of the root-mean-square of successive differences of the interbeat intervals during sleep. For RHR and HRV, higher weights were assigned to periods with a higher probability of slow wave sleep and periods closer to the end of the sleep. Respiratory rate is calculated as the median of respirations per minute calculated via the interbeat intervals throughout sleep. Heart rate zones (HRZs) were determined as a percentage of the participant’s estimated maximal heart rate, where zone 0<50% of maximal heart rate, zone 1=50%‐60%, zone 2=60%‐70%, zone 3=70%‐80%, zone 4=80%‐90%, and zone 5=90%‐100%. Physical activity measures included percent time spent in HRZ 1‐5, summated HRZs, and physical activity level (total energy expenditure divided by resting energy expenditure). Summated HRZ was calculated as the time spent in each HRZ, in minutes, multiplied by the corresponding factor for each zone and then summated. Physical activity level was calculated as the total calories expended divided by resting energy expenditure. Cardiovascular, respiratory, and sleep measures derived throughout sleep have been validated against gold standard polysomnography measures and have been found to have a low degree of bias and low precision errors (eg, <20 min bias and precision errors for sleep duration; 0.7 beats per min bias for heart rate; 4.7 ms bias for HRV; and 1.8% bias for respiratory rate) [23-25]. More details on the methods for calculating wearable-derived metrics can be found in Multimedia Appendix 2.

Survey responses from individuals older than 21 years of age were considered for this study. Each survey response from the participant over the 13-month period was paired with the wearable data from the 14 or 28 days preceding the survey response. Furthermore, 14 days of wearable data leading up to the response were used for the PHQ2 and GAD2 as the questions required participants to reflect on how they felt over the previous 2 weeks; 28 days of wearable data leading up to the response were used for the PSS as the questions required participants to reflect on how they felt over the previous month. Except for the measures that pertained to variability, the daily measures from the time leading up to completing a survey were averaged via the mean. For measures pertaining to variability, except for sleep consistency, the sample SD of the daily measures over the relevant period of time was used. Periods during which a user had recorded fewer than 7 days of wearable data in either the 14- or 28-day period leading up to completing their survey were excluded from analysis. Participants were included if they selected man or woman when asked to self-identify their gender. Extreme values were clipped at 17 and 40 kg/m² for BMI and at 70 years for age. BMI was also examined by binning into classification ranges: underweight (<18.5 kg/m²), healthy weight (≥18.5 kg/m² and <25 kg/m²), overweight (≥25 kg/m² and <30 kg/m²), and obese (≥30 kg/m²). As this investigation included participants from both the northern and southern hemispheres, which experience opposite seasons for a given date, we adjusted the submit month of participants from the southern hemisphere by shifting each submit month to 6 months forward. For example, responses submitted in June from participants in the southern hemisphere were adjusted to December to align the seasons of the northern and southern hemispheres.

A subsample of individuals was used to understand the intrapersonal relationships and the relationship between physical activity, estimated via HRZs, and mental health status. Eligibility for this subsample required that participants responded to the survey at least 8 times and recorded more than six 24-hour periods containing at least 1000 minutes of data in the time frame leading up to each of the survey responses.

Generalized mixed linear models were used to examine the relationships between mental health and age, gender, BMI, and month of submission. In these models, the mental health outcome was modeled as the dependent variable; age, gender, BMI, and month of submission were treated as the independent variables; and each participant was treated as a random effect. Natural cubic splines were used for BMI and month of submission due to the nonlinear relationships these variables displayed with mental health. Estimated marginal means for age, gender, BMI, and month of submission were extracted from the models to provide estimates and variances for plotting. To examine the effects of different BMI categories and seasons on mental health, pairwise comparisons were made using the emmeans package in R (R Core Team).

Generalized mixed linear models were also used to examine the relationships between mental health responses and wearable-derived metrics. In these models, mental health outcomes were modeled as dependent variables, and the wearable-derived metrics as the independent variable. Age, gender, BMI, and the month of submission were used as covariates in the model, and each participant was treated as a random effect. For intrapersonal associations, wearable-derived metrics were person-mean-centered. Natural cubic splines were used for BMI and month of submission due to the nonlinear relationships these variables displayed with mental health and wearable-derived metrics. The raw and standardized coefficients (β) from the models were extracted to describe the relationship between wearable metrics and mental health.

To examine the temporal relationships between mental health outcomes and physiological variables, we conducted a lagged analysis using structural equation modeling via the lavaan package in R. For these models, both mental health outcomes and physiological variables were centered within each participant by subtracting their respective within-person means. Lagged variables were generated for both mental health outcomes and physiological variables by using the immediate preceding mean person-centered observation. We accounted for the nested structure due to repeated measures within participants by using the built-in clustering functionality. To flexibly model time of year effects, natural spline transformations with 3 degrees of freedom were applied to the submission month. Demographic variables—such as age, BMI, and gender—were assumed to be invariant over time and modeled as such. BMI and age were further transformed using natural splines (with 3 degrees of freedom) to capture potential nonlinear effects in the relationships with the outcomes. Autoregressive effects were modeled by regressing each outcome on its own lagged value. Contemporaneous covariance was addressed by allowing for residual correlations between the mental health variable and physiological variable at the same time point. Reciprocal influences were examined by regressing the mental health variable on the lagged physiological variable and vice versa, enabling an exploration of bidirectional effects. Cross-lagged paths were estimated within the structural equation model framework, with standardized parameter estimates extracted to facilitate interpretation of temporal relationships (simplified diagram of cross-lag design in Multimedia Appendix 3).

The significance threshold for all analyses was set to 0.001.

Participants consented to their anonymized data being used for research purposes. Since data were not identifiable and were stored on a secure server, this study was deemed exempt from institutional review board oversight by Advarra’s Institutional Review Board (Columbia, Maryland). Participants did not receive compensation for their participation in this study.

A total of 181,574 individuals were considered for eligibility, and, after excluding due to either gender and data availability, 172,283 individuals were deemed eligible for the PSS analysis and 170,320 were considered eligible for the PHQ2, GAD2, and combined score analyses. Among the subset of individuals included in mental health and physical activity association analyses, 3197 participants were included for the PSS, and 3196 were included participants for the PHQ2, GAD2, and combination score analysis (Figure 1). The number of days included in each analysis and demographic information for each subgroup are provided in Table 1.

Self-reported mental health scores decreased with age (indicating better mental health) across all mental health outcomes (Figure (a) in Multimedia Appendix 4; PHQ2: P<.001, standardized β coefficient=−0.019; GAD2: P<.001, standardized β coefficient=−0.023; PSS: P<.001, standardized β coefficient=−0.019; combined: P<.001, standardized β coefficient=−0.06). On average, males reported better mental health on their assessments than females (Figure (a) in Multimedia Appendix 4). Mental health scores shared a nonlinear relationship with BMI, whereby mental health was worse at the upper (ie, obese) and lower (ie, underweight) extremes as compared with a healthy or overweight BMI (Figure (b) in Multimedia Appendix 4; all P<.001). Mental health scores also displayed a seasonal pattern, whereby individuals reported worse depression in the Fall as compared with all other seasons and worse stress in the Fall as compared with Spring and Winter (Figure (c) in Multimedia Appendix 4; all P<.001). Pairwise comparisons among BMI groups and seasons can be found in Multimedia Appendix 5. Due to the associations observed between demographics and seasonality with mental health outcomes, we incorporate demographics and seasonality as covariates in our models examining the relationships between physiological variables and mental health outcomes.

We conducted a comprehensive longitudinal analysis on the inter- and intrapersonal relationships between wearable-derived metrics and mental health outcomes. We find that responses to mental health surveys differ depending on age, gender, BMI, and time of year. After controlling for these demographic and seasonal variables, better mental health outcomes were found in individuals who had (1) longer sleep durations; (2) higher sleep efficiencies and HRV_avs; (3) lower variation in their sleep patterns; (4) lower RHRs, HRV_cvs, and respiratory rates; (5) and spent more time being physically active. Regarding intrapersonal observations, better mental health outcomes coincided with increases in sleep efficiency, sleep consistency, HRV_av, and physical activity and decreases in RHR, respiratory rate, and HRV_cv. Notably, changes in sleep duration displayed distinct associations with mental health outcomes when examined intra- or interpersonally, whereby an increase in sleep duration within an individual was associated with worse depression scores yet better anxiety and stress scores. Overall, this information extends our understanding of the intricate connections between physiological data and mental well-being, providing valuable insights for health care professionals, researchers, and policy makers aiming to enhance mental health interventions and personalized care strategies.

Mental health exhibited seasonal variation, though the clinical meaningfulness of these fluctuations is uncertain. Some individuals report seasonal changes in mood that surpass clinically significant thresholds, such as those observed in seasonal affective disorder [26,27]. Several factors may contribute to these seasonal mood variations, including reduced sun exposure leading to changes in hormone levels (eg, melatonin and serotonin), as well as alterations in physical activity or sleep patterns [28,29]. In addition, it is noteworthy that seasonal variations have also been observed in cardiovascular metrics such as RHR, HRV, and even blood pressure [30]. Whether these seasonal changes in physiology or mood can be tempered by lifestyle modification remains uncertain but is an intriguing area of further exploration [31,32].

Individuals who slept more and had lower levels of variation in their sleep metrics displayed better scores on mental health assessments. This finding was corroborated by within-person analyses showing that increases in both sleep duration and consistency are linked with lower stress and anxiety levels. However, an interesting divergence in within- and between-person findings appears when we examine depression: while greater sleep duration was associated with lower scores of depression scores between individuals, higher depressive scores were associated with less sleep within person. One possible explanation for the differences observed in between- and within-person analyses may be due to lower amounts of sleep increasing the risk of depression [33], which would be reflected in interpersonal observations. Conversely, increases in sleep durations have been shown to be symptoms of depression [34], potentially when depression is incident with higher levels of inflammation [35], which would be reflected in intrapersonal observations. Overall, the opposing between- and within-person associations between sleep duration and depression demonstrate the importance of measuring the change in biometrics as an addition to single or sporadic measures of biometrics.

In both inter- and intrapersonal analyses, higher and more stable HRV and lower RHR values were associated with more favorable mental health profiles. The coincident changes in mental state and cardiovascular measures may stem from their interactions with the autonomic nervous system. It has been shown that individuals with higher levels of depression, anxiety, or stress tend to exhibit increased sympathetic output or diminished parasympathetic activity [36]. Regardless of the underlying mechanism, the link between RHR and mental health outcomes is well-established. For instance, a large-scale longitudinal study of over 1 million adolescent men found that a higher RHR is linked with an increased risk of psychiatric disorders [37]. Similarly, reduced HRV has been documented in psychiatric conditions such as depression, substance abuse, anxiety, and schizophrenia [38]. Although less extensively studied, day-to-day fluctuations in HRV, indexed by HRV_cv, represent an important measure of autonomic balance with both physiological and psychological implications [39]. The association of higher HRV_av or lower RHR and HRV_cv with reduced depression, anxiety, and stress levels in our study contributes to a growing body of literature that supports the utility of tracking the cardiovascular system with wearables, as changes to these cardiorespiratory metrics may coincide with changes in mental health status.

Our results suggest that the relationship between exercise and mental health may depend upon exercise intensity. For example, the percentage of time spent in the highest intensity of exercise, zone 5, is only associated with better mental health outcomes for intrapersonal associations with depression, whereas total time spent being physically active (as determined by physical activity levels) is consistently associated with better mental health outcomes. Meta-analyses looking into the effect of interventional high-intensity exercise on mental health indicate that high-intensity exercise generally improves mental well-being [40-42]; however, too much high-intensity exercise without adequate recovery can lead to overtraining, and overtraining has been shown to be detrimental to mental health [43,44]. Meanwhile, we consistently observed that lower intensity exercise or general physical activity associates with improved mental health outcomes, which aligns with findings from previous investigations [45]. Interestingly, when comparing mental health benefits of moderate-to-vigorous continuous training (at an intensity likely to be well below HRZ 5) versus high-intensity interval training (HIIT; at an intensity likely to reach HRZ 5), moderate-to-vigorous continuous training led to significant improvements in positive affect and mental well-being over control and HIIT; furthermore, HIIT failed to improve mental health either from baseline or compared with control conditions [46]. Nevertheless, future research is warranted to better understand the potential dose-effect of high-intensity exercise on mental health.

The significant associations observed in our between- and within-person analyses but not in our cross-lagged models may be due to the temporal dynamics of physiological and psychological interactions. The mental health questionnaires (GAD2, PHQ2, and PSS) assessed symptoms over the previous 2‐4 weeks, and we aligned physiological variables with these same multiweek windows. However, psychological and physiological processes are often intertwined on much shorter timescales. Empirical evidence suggests that stress and related physiological responses (such as sleep and cardiovascular changes) are tightly linked in the short term, typically over days or even hours [47-50]. For instance, acute psychosocial stress can immediately disrupt sleep (transient “acute” insomnia), an effect that generally resolves once the stressor is removed [51]. Conversely, a single night of poor sleep can heighten stress levels and autonomic arousal (eg, elevated heart rate) the following day, whereas a good night’s sleep has the opposite effect [47]. These observations underscore the reciprocal, short-lived nature of stress-sleep interactions reported in the literature [52]. Because our analyses spanned up to 4 weeks – a time frame designed to capture recent stress exposure – any acute bidirectional effects would presumably be contained within this assessment window and potentially explain why we failed to observe more temporal relationships. Future studies using high-frequency, real-time assessments may provide further insight into the dynamic nature of these associations and help disentangle acute versus chronic effects.

Our findings should be interpreted in the context of this investigation’s limitations. First, selection bias may be present [53], as individuals who can afford commercial wearable devices likely represent a higher socioeconomic status. In addition, filtering for compliance within the HRZ analyses may select for a subgroup that differs from the general population. To mitigate the impact of these potential biases, we leveraged a large sample size and conducted within-person analyses. Second, our findings may be subject to recall bias since participants were asked to reflect on their mental health over the previous 2 weeks or month at the time of completing surveys. To address this, future studies could incorporate ecological momentary assessments, allowing for real-time data collection to minimize recall inaccuracies. Third, events or environmental factors not measured in our study may influence the relationship between mental health and wearable data. To partly mitigate this limitation, we leveraged data from a large and diverse sample spanning multiple countries, potentially reducing the impact of localized events or environmental factors. Fourth, we did not control for menstrual phase when analyzing cardiovascular or survey data in female participants, which may introduce variability. Future studies could incorporate menstrual cycle tracking to control for physiological variations related to menstrual phases. Fifth, this investigation specifically measured stress, anxiety, and depression using the PSS, GAD2, and PHQ2 scales and did not examine other mental health disorders such as bipolar disorder or post-traumatic stress disorder. We refrained from including additional diagnostic surveys to avoid increasing participant burden and reducing study retention. However, future research may benefit from incorporating broader diagnostic measures or clinical evaluations to more comprehensively capture mental health outcomes.

Leveraging a large sample size, longitudinal design, and continuous second-by-second data, this study provides valuable insights into the relationships between wearable-derived physiological metrics and mental health outcomes. The associations observed across multiple biometric indicators suggest that wearables hold promise as complementary tools in mental health monitoring. As our understanding of the dynamic interplay between physiological and psychological states continues to evolve, this research adds to the growing body of evidence that wearables can help elucidate the intricate relationship between mental states and physiological processes. Future research should further explore how wearables can be leveraged in real-time assessments and adaptive interventions to enhance early detection, prevention, and treatment of mental health conditions.