Psychological stress is associated with increased risk of several forms of cancer [1], musculoskeletal diseases [2], periodontal diseases [3], type 2 diabetes mellitus [4], stroke [5], cardiovascular disease [6], and recurrent cardiovascular disease [7]. In an occupational setting, psychosocial risk factors such as high job demands are estimated to increase the risk of stress-related diseases (eg, burnout) by 60%-90% [8]. Occupational stress can therefore cause absenteeism, organizational dysfunction, and decreased productivity, and it has a large economic burden [9].

Stress occurs when the brain subconsciously appraises a demand as threatening because of a lack of resources to cope with it [10]. This threat appraisal that we refer to as stress is sometimes referred to as distress, whereas demands for which sufficient coping resources are available are appraised as a challenge or as eustress. Therefore, stress can be seen as a psychological state that is the result of a divergence between demands on an individual and the individual’s perceived capacity to cope with them. Stress causes an imbalance in the body’s biological equilibrium (homeostasis), which requires a neural, neuroendocrine, and neuroendocrine-immune adaptation to restore it (allostasis) [11,12]. Although acute stress can have negative effects, it is particularly the cumulative wear and tear on bodily systems (allostatic load) caused by excessive stress or inefficient management of the systems that promote adaptation that is detrimental to long-term health and well-being [13]. In addition, lifestyle-related factors such as obesity, sleep, and substance abuse can also contribute to allostatic load [14]. Allostatic load is therefore considered a measure of the cumulative biological burden on health [15].

To complement this biological and neuroendocrinological perspective on the negative long-term health effects of stress and provide a framework for how short-term spillover effects of stress accumulate the need for a recovery concept be introduced [16]. A need for recovery arises when an individual has problems using resources to adaptively cope with demands that induce stress [17]. Need for recovery is a conscious emotional state that is related to the temporal depletion of resources following effort to meet demands and is characterized by feelings of mental exhaustion [18]. As the availability of resources is assessed during appraisal and the use of resources may be needed during coping, the Conservation of Resources Theory states that an initial loss of resources can lead to a negative feedback loop that increases one's vulnerability to stress [19]. Such a loss spiral may become even more distinct if stress negatively impacts the recovery process itself, for instance, by negatively impacting sleep quality [20] and psychological detachment [21]. Resilience, which can be defined as the process of positively adapting to adverse events [22], is a term describing this process from a positive perspective. During a resilient process, the aforementioned loss spiral is prevented by using resources to adaptively cope with demands and stress to limit long-term strain and its related negative consequences on health and well-being from developing [23]. Resilience is therefore an ongoing process that influences the extent to which adverse events that occur on a small timescale have an impact on mid- to long-term health outcomes.

A challenge for resilience research that focuses on resilience-related associations on timescales is that it requires continuous data collection, making it relatively labor-intensive for participants to do so. Over the past decade, the emergence of smartphones and wearable sensor technologies has enabled the easy and unobtrusive measurement of physiological and psychological data related to an individual’s resilience [24]. A promising example of such a metric is heart rate variability (HRV), which refers to the variation in interbeat intervals of the heartbeats [25]. HRV is a plausible, noninvasive, and easily applicable biomarker for resilience that may serve as a global index of an individual’s flexibility and adaptability to stressors [26,27]. HRV is negatively correlated with allostatic load, illustrating its use as an overall health risk indicator [28]. Stress is also known to decrease HRV, particularly with reduced parasympathetic activation [29-31]. Although an acute decline in HRV may be indicative of increased acute stress levels, HRV can remain lowered during rest and sleep after stress or mental exhaustion [32-35]. Conversely, having a lower trait resting HRV has been linked to increased sensitivity to stress via appraisal when faced with demands [36] and to suboptimal emotion regulation that may result in mental exhaustion [37,38]. Therefore, resting HRV can be seen as a physiological resource that is addressed during the appraisal of demands and coping with stress. Therefore, resting HRV can be hypothesized to have a buffering effect on the positive associations between demands and stress, as well as between stress and mental exhaustion. These two hypothesized buffering effects are depicted as circles 1 and 2 in Figure 1, which represent the conceptual model for this study and were based on a previous publication [39]. The model is based on the Transactional Model of Stress and Coping [10], the Job Demands-Resources Model of Burnout [40], the Effort-Recovery Model [17], and the Conservation of Resources Theory [19]. In short, it depicts that demands are appraised as stress when resources are low, that stress leads to mental exhaustion when resources to cope with the demands are lacking, and that mental exhaustion limits resources to deal with future demands, unless there are sufficient recovery opportunities. In this study, HRV is the resource of interest, whereas sleep, operationalized as total sleep time (TST), represents the model’s recovery process.

Besides resting HRV, sleep is also a relevant potential indicator for the accumulation of the negative consequences of stress and predictor of spillover need for recovery. In the literature, stress has been consistently shown to decrease slow-wave sleep, rapid eye movement sleep, and sleep efficiency, as well as to increase the number of awakenings that may impact the overall sleep duration [20]. Therefore, stress can be hypothesized to negatively affect the TST, which is the total time during a sleep episode in which one was not awake (Figure 1; hypothesis 3). In contrast, sleep has important homeostatic functions that are essential during recovery from both physiological and psychological strains [41]. Sleep deprivation therefore has been linked to an increase in allostatic load [42] and has been linked to decreased HRV in some studies [43,44]. As mental exhaustion may result in decreased resting HRV [34,35] and sleep is an essential aspect of the recovery process, TST can be hypothesized to buffer against the negative association between mental exhaustion and resting HRV (Figure 1; hypothesis 4).

HRV measurement is regularly used as a biofeedback tool in mobile health (mHealth) interventions that target acute stress relief [45-47] but may also be useful for interventions that aim to provide users with feedback on their resilience over a longer timeframe. A recent literature review confirmed that HRV has potential as a biomarker for resilience but suggested that more longitudinal studies are needed that use wearable sensors to observe HRV in a naturalistic context of real-life and associate it with resilience-related outcomes, as most of the evidence is based on cross-sectional population studies [27]. Therefore, this study longitudinally assessed the aforementioned hypotheses in a free-living context using consumer-available wearable sensors. Exploring these will provide insight into the potential causal pathways of the within-day accumulation of the negative consequences of stress. Gained insights may therefore be beneficial to the future development of (automated) resilience interventions that target the prevention of stress-related health problems.

Students in applied psychology, social work, and physiotherapy who were about to start their first full-time internship were invited to participate via a message on the school’s digital learning environment and email. This population was anticipated to be at risk of experiencing stress because of the potentially stressful nature of these internships, as well as the fact that this was the first internship in the participants’ curriculum. A maximum of 15 participants could be simultaneously recruited because of the availability of materials. The recruitment and data collection processes were therefore divided into two waves that started in September 2018 and September 2019, respectively. The students were sent an email with an information letter that described the goal of the study, a description of the measurement protocol, and management of the collected data; the email also stated that participation would be unrelated to their internship or educational progress, that participation would occur on a voluntary basis, and that they could stop at any time without negative consequences. Some of the researchers were employed by the university in which the students were enrolled but had no other associations with the invited students (eg, via education). The participants provided written informed consent before participation. Participants who collected complete data on at least 84 days during the formal participation period were rewarded with a €25 (US $27.50) gift voucher to facilitate recruitment and optimize adherence during participation. This reward threshold represents an adherence of at least 80% over a data collection period of 15 weeks (105 days). The threshold was solely used as a cutoff point for the reward and not for statistical analyses.

Participants were assisted in installing the required apps on their smartphones and were instructed on how to use the devices used for data collection. The data collection period started immediately after the measurement instructions, after which participants collected data for 15 weeks. Some participants completed additional daily measurements on a voluntary basis until their appointment was planned to return the used materials. At this appointment, an additional 20-minute conversation was held about how they experienced the daily measurements and to learn about potential improvements for future studies. During the study, anonymized user accounts were used for the applied consumer-available wearables in order to protect the participants’ privacy on the companies’ cloud servers, before being exported and deleted by the researchers after completing their participation.

All data management and analyses were performed in RStudio [68] and R [69].

A total of 26 participants were recruited for this study. The participants were predominantly women (n=24). Most participants studied applied psychology (n=19), followed by social work (n=6) and physiotherapy (n=1). The participants were aged between 19.2 and 33.2 years (median 22.6). The participants collected TST data on 2129 days (per participant range 10-119; median 94), 1731 morning EMA questionnaires (range 5-109; median 74), 1653 evening EMA questionnaires (range 7-111; median 73), and HRV data on 1443 days (range 6-115; median 53). In total, for 1004 of the 2379 days (42.2%) on which a participant collected data, the participant collected complete data containing all required TST, HRV, and EMA data. The descriptive statistics for and intercorrelations between the main and control variables are presented in Table 1. Three participants did not complete the full (105 days) measurement period because they lost motivation for the daily measurements, and one participant stopped the daily measurements because of skin rash related to wearing the Fitbit. All three participants who did not complete the full measurement period contributed daily measurements and were thus still included in the analyses. During the exit conversations, several participants stated that they found it difficult to adhere to the HRV measurement, because the need to apply a moistened chest strap and lay still for 2 minutes after awakening while they wanted to continue with their day was inconvenient. Missing Fitbit data were primarily ascribed to forgetting to charge it, particularly when participants were away from home. Finally, participants mostly attributed missing EMA data to simply forget to act on the smartphone notification as they were busy at that time.

A two-step hierarchical multiple regression model explaining stress scores was developed (Table 2). After controlling for MVPA, sedentary time, and previous-night TST, demands were positively associated (P<.001) with stress. In addition, the interaction effect of demands and resting HRV significantly (P=.044) buffered against this positive association. This means that participants tended to report higher stress scores on days that they also considered to be more demanding, but this relationship was weaker on days where the participant woke up with a relatively high resting HRV. The positive association between demands and stress, as well as the buffering effect of resting HRV, confirms hypothesis one. Furthermore, the control variable MVPA was positively associated with stress (P=.044), which means that participants reported higher stress scores on days where they were more physically active. In the control variable model, TST was a negative predictor of stress (P=.03), but this effect was no longer significant in the full model. The control variable model of analysis explained 2.0% of the within-subject variance in the daily stress scores, whereas the full model had an explained variance of 21.7%, which is a significant increase from the control variable model.

A two-step hierarchical multiple regression model explaining mental exhaustion scores was developed (Table 3). After controlling for MVPA, sedentary time, and previous-night TST, stress was positively associated (P<.001) with mental exhaustion. In addition, the interaction effect of stress and resting HRV significantly (P=.029) buffered against this positive association. This means that participants tended to report higher mental exhaustion scores on days that they were also considered stressful, but this relationship was weaker on days where the participant woke up with a relatively high resting HRV. The positive association between stress and mental exhaustion, as well as the buffering effect of resting HRV confirm hypothesis two. In the control variable model, MVPA was also positively associated with mental exhaustion (P=.017), but this effect was no longer significant in the full model. The control variable model of analysis two explains 1.4% of the within-subject variance in the daily mental exhaustion scores, whereas the full model has an explained variance of 31.6%, which is a significant increase from the control variable model.

A two-step hierarchical multiple regression model explaining nighttime TST was developed (Table 4). After controlling for previous-day MVPA, sedentary time, and alcohol consumption, previous-day stress did not predict TST, unlike our expectation based on hypothesis 3. However, the control variables previous-day MVPA and alcohol consumption were negatively associated with TST, whereas sedentary time was positively associated with TST. This means that participants had a lower TST on days where they were relatively physically active, consumed alcohol, and had limited sedentary time. The control variable model of analysis three explains 3.8% of the within-subject variance in TST, whereas the full model has an explained variance of 4.6%, which is not a statistically significant increase from the control variable model.

A two-step hierarchical multiple regression model explaining resting HRV was developed (Table 5). After controlling for previous-day MVPA, sedentary time, and alcohol consumption, previous-evening mental exhaustion negatively predicted (P<.001) resting HRV, but previous-night TST did not buffer against this negative association. Therefore, these results partially support hypothesis four. Among the control variables, previous-day alcohol consumption negatively predicted resting HRV (P<.001). The control variable model explained 2.3% of the within-subject variance in resting HRV, whereas the full model had an explained variance of 3.6%, which was not a statistically significant increase from the control variable model.

This study aimed to test the hypotheses that (1) demands are positively associated with stress and resting HRV buffers against this association, (2) stress is positively associated with mental exhaustion and resting HRV buffers against this association, (3) stress negatively impacts subsequent-night TST, and (4) previous-evening mental exhaustion negatively impacts resting HRV, while previous-night TST buffers against this association. By assessing these associations based on longitudinal data that were collected using consumer-available wearables and smartphone apps in a free-living context, this study provides insight into the potential pathways of the within-day accumulation of the negative consequences of stress. The results of this study support hypotheses one and two and partially support hypothesis four.

As hypothesized, having a high resting HRV buffered against the positive associations between demands and stress (hypothesis 1), as well as between stress and mental exhaustion (hypothesis 2). Similarly, mental exhaustion negatively predicted resting HRV, as expected (hypothesis 4). These findings suggest that waking up with a relatively high intraindividual resting HRV decreases an individual’s sensitivity to stress when faced with demands, as well as the likelihood of being mentally exhausted during a stressful day. In addition, as the accumulation of mental exhaustion negatively impacts an individual’s resting HRV, an increase in mental exhaustion negatively impacts this (psycho) physiological resource and thus potentially creates a negative feedback loop that can lead to a loss spiral. These results therefore confirm our hypothesis that a decline in resting HRV is indicative of the accumulation of the negative consequences of stress, as well as the continued accumulation of negative consequences of stress. Therefore, resting HRV can be seen as a biomarker for or an index of resilience, where a decline in resting HRV may signal that buildup of allostatic load is present and suggests that the individual’s readiness to face new demands may at least be temporarily decreased.

As highlighted in a recent literature review, most studies to date investigating the role of HRV as an index for resilience have a cross-sectional nature and assess relationships at the between-subject level [27]. To our knowledge, this study is the first to apply a nested longitudinal design and assess the potential of resting HRV as an index of resilience to stress on a within-subject level. Previous studies that investigated between-subject differences identified similar relationships between resting HRV and mental health outcomes. For instance, a recent study with school teachers concluded that 48-hour trait HRV buffered the effect of emotional demands on exhaustion [75]. Another recent study cross-sectionally assessed a population of young female adults and found that having a high resting HRV buffered against the positive association between emotion regulation difficulties and depressive symptoms [76]. Resting HRV has also been reported to buffer against the negative effects of chronic stress on sleep quality, which in turn is related to greater depressive symptoms [77]. Finally, high stress-induced HRV was shown to buffer against the negative effect of hostility on cortisol sensitivity [78]. Therefore, the within-subject findings of this study align with previous studies that also reported favorable between-subject effects of resting HRV on diverse mental health outcomes.

Contrary to our hypothesis, stress did not negatively affect TST (hypothesis 3). The absence of a negative association between stress and TST conflicts with previous literature that consistently links experimental stress to decreased slow-wave sleep, rapid eye movement sleep, sleep efficiency, and increases in awakenings [20]. A possible explanation for this could be the difference in context, as those studies examined the influence of experimental stress on polysomnographically measured sleep, whereas this study investigated daytime stressors and TST in a natural free-living context. Because of the increasing capabilities and performance of consumer wearables to measure sleep and the resulting rise in the use of consumer wearables in sleep research, future studies may increase insights into the potential relationship between daily stressors and TST in a natural free-living context [79].

TST also did not buffer against the negative association between mental exhaustion and resting HRV, as expected (hypothesis 4). This expectation was based on the rationale that sleep has important homeostatic functions that are essential during recovery from strain [41], and that sleep deprivation has been linked to an increase in allostatic load [42] and a decrease in HRV [43,44]. As mental exhaustion was measured during the evening and resting HRV during the morning, we expected TST to potentially have a buffering effect, meaning that the negative impact of mental exhaustion on resting HRV would be smaller if the participant slept well that night. This buffering effect was not present in these findings, but TST was also not positively associated with resting HRV, as might be expected based on the aforementioned literature. A possible explanation for this could be that the relationship between sleep deprivation and HRV in previous literature seems to be particularly present in studies assessing a longer sleep deprivation period [80], which might suggest that the nuanced day-to-day differences in TST are too small to significantly impact the resting HRV. Future studies investigating the impact of TST on resting HRV or the recovery from strain in a natural free-living context in which such long periods of sleep deprivation are relatively uncommon, assessing the impact of multi-day trends in TST might help increase insight on this topic.

The effects of most of the control variables that were significantly associated with the outcomes of the four analyses were as expected, but some of the effects seem to conflict with previous literature. For instance, MVPA was negatively associated with TST, but a recent study found a positive association between MVPA and TST [60]. Similarly, sedentary time was positively associated with TST in this study, whereas a negative association with TST was reported in another recent study assessing obese adults [62]. A possible explanation can be found in the reported significant correlations between MVPA, sedentary time, and alcohol consumption (Table 1). This young student population spends part of their leisure time enjoying the local nightlife, in which dancing and alcohol consumption are common. It is therefore possible that this will have caused the low sedentary time, high MVPA, and alcohol consumption that were associated with low TST.

The strengths of this study are its longitudinal design and large sample of nested observations, optimizing the within-subject variance. Moreover, the use of consumer-friendly wearable sensors and smartphone apps allowed for relatively unobtrusive monitoring in a free daily living context, optimizing the generalizability for similar settings. A limitation of this study was the need to apply relatively coarse algorithmic artifact correction and rule-based outlier filtering during HRV data management. Because of the choice to use a consumer-available sensor and app for long-term daily measurements in free-living conditions, electrogram-level data were unavailable, and it was impossible to verify if participants performed the measurements as instructed. As the applied algorithmic artifact removal method can only filter out interbeat interval artifacts within an HRV measurement, it has no rules to decide whether an observation should be removed altogether, filtering out extreme within-subject RMSSD outliers was necessary. Furthermore, algorithm-based artifact correction was preferred over manually adjusting interbeat interval artifacts to make the findings of this study applicable to the context of an automated resilience intervention that does not rely on human interference during data management. In addition, the use of single-item scales in the evening EMA questionnaire forms a limitation of comprehensiveness at which the concepts can be measured. Therefore, validated single-item scales or items with the most favorable psychometric properties in existing validated scales were used where available [39]. As single-item scales have consistently been found to be valid measures for diverse concepts in comparison to full scales [81-84] and have become common good in EMA research, the applied EMA methods can still be considered appropriate. The participants also received some feedback on their sleep, physical activity, and HRV because of the use of the consumer-available Fitbit and Elite HRV apps, which might have influenced their behavior. Nevertheless, any such influence was not considered a problem because this study observes the natural relationship between several variables and does not reflect on the behaviors themselves. Finally, only 3.6% of the within-subject variance in resting HRV could be explained, and the buffering effect of resting HRV was relatively modest.

The HRV-related results can be generalized to young and employed female adults who track their resting HRV upon awakening. As 92% (24/26) of the participants were female and HRV can be related to menstrual cycle changes [85], further research on young males is necessary to improve the generalizability of these findings to young adults regardless of gender. As the resting HRV was measured upon awakening in this study, the influence of a phenomenon called the cortisol awakening response (CAR) might have played a role. Upon awakening, cortisol levels start to increase and peak approximately 30-45 minutes thereafter because of the CAR, where 1%-3.6% of its variance can be explained by psychosocial factors [86]. Although the CAR is associated with postawakening changes in HRV, these changes appear to be unrelated to perceived stress and measures of emotion regulation [87]. Therefore, it is possible that measuring HRV during sleep could yield similar results. An advantage of measuring resting HRV during sleep is that participants would not need to apply a moistened chest strap and lay still in a supine position upon awakening to collect their resting HRV data. As multiple participants described that this procedure negatively impacted their adherence to the measurement protocol, unobtrusively measuring the resting HRV during sleep might improve adherence and thus increase statistical power. Future research is needed to confirm whether resting HRV during sleep can be used to yield similar results.

To our knowledge, this study is the first to report a significant within-subject buffering effect of resting HRV on the positive associations between demands and stress, as well as between stress and mental exhaustion and a negative association between mental exhaustion and resting HRV. Replication of these findings in future studies is needed. As the combined findings form a feedback loop, it is possible that multi-day trends in resting HRV could be linked to longitudinal mental health outcomes in future studies. Furthermore, exploring the use of time series analysis to create within-subject models in which multi-day trend data are used to assess the daily outcomes could potentially improve the accuracy of the presented models.

Future studies are advised to use passive monitoring techniques that require little to no user attention whenever possible to improve participant adherence and optimize statistical power.

If the findings of this study can indeed be replicated and expanded upon, it would show that longitudinally monitoring resting HRV as a biomarker of or index for resilience may be useful in the context of prevention. In this context, a structural increase or decline in resting HRV could provide an early warning signal that a positive or negative feedback loop is formed. When used in a consumer wearable–based automated resilience intervention, these signals can be used to prompt user feedback. For instance, users could be rewarded when a positive feedback loop is recognized or suggested to perform cognitive behavioral therapy–based self-reflection exercises or relaxation techniques when a negative feedback loop occurs. Such an automated resilience intervention that unobtrusively monitors the user’s resting HRV for the early recognition of (un)favorable feedback loops and generation of just-in-time feedback may therefore limit the buildup of allostatic load and improve long-term health outcomes.