Assessing Stress Level Scores Against Wearables‐Driven Physiological Measurements

ABSTRACT

Daily stressors elicit physiological and mental responses impacting health, cognition, and behaviour. Accurately assessing psychological stress responses in natural settings remains challenging despite extensive research, though wrist‐worn devices have the potential to address this gap through remote data collection. The Garmin fitness tracker provides a stress score largely based on HRV which must be validated prior to use in research. This study aimed to (1) predict psychological self‐reported stress from physiological measurements and Garmin calculated stress score, and (2) assess the stress score given by the Garmin Vivosmart 4 against HR and HRV from ECG recordings derived by the Polar H10 chest strap. A pilot study of 29 participants was conducted, followed by power calculations and preregistration of the main study which included 60 participants. Data were collected simultaneously from both Garmin and Polar device during a laboratory session of restful and mental‐stress‐inducing tasks. Garmin's stress score, mean HR, SD2/SD1, and HF power exhibited significant differences between stress and rest conditions. Moreover, Garmin's stress score correlated significantly with HR, RMSSD, and SD2/SD1. However, out of our physiological measurements, heart rate showed the strongest association with self‐reported stress, while the Garmin stress score demonstrated only marginal predictive value for subjective stress experience. Our findings also suggest that physiological responses to mental stress were influenced by sex and tonic HRV. The study suggests that the GSS, although even better heart rate, are indicative of mental stress. Garmin, with its accessibility and noninvasive nature, measures both heart rate and consumer health score (stress), promising widespread utilisation in various research domains.

1. Introduction

In our daily lives, we encounter a multitude of triggers that elicit various forms of stress responses related to challenging or threatening situations (Almeida 2005). These include physiological and mental reactions that play a pivotal role in shaping our health, cognition, and behaviour (Schneiderman et al. 2005) (Lupien et al. 2009). Mental stress has emerged as a pressing concern in contemporary society, exerting a profound impact on both physical and psychological well‐being. The term 'stress' encompasses the body's response to internal or external conditions (i.e., stressors) that disrupt homoeostasis (Chrousos 2009), triggering physiological and adverse psychological changes.

1.1. Psychological Stress

Monitoring psychological stress in the laboratory is a critical component of research aiming to understand stress responses and their consequences for health, behaviour, and cognition. Standard laboratory paradigms, such as the non‐verbal Montreal Imaging Stress Task (MIST), reliably induce acute stress through evaluative tasks that mirror real‐life challenges, allowing for consistent measurement across participants (Dedovic et al. 2005). Researchers employ a multi‐method approach to assess stress, combining physiological indicators (such as heart rate, blood pressure, skin conductance, and respiratory rate), and validated self‐report questionnaires like the Perceived Stress Scale (PSS) (Cohen et al. 1983). These complementary measures provide a comprehensive understanding of the interplay between subjective experiences and objective biological responses, facilitating the study of both acute and chronic psychological stress.

1.2. Physiological Stress Responses

Physiological stress responses involves the activation of two major pathways: the Sympathetic‐Adrenal‐Medullary (SAM) axis, which releases noradrenaline and instigates the ‘fight or flight’ response, and the Hypothalamus–Pituitary–Adrenal (HPA) axis, responsible for the release of glucocorticoids, primarily cortisol in humans (Giannakakis et al. 2019). Physiological manifestations of stress include heightened sympathetic nervous system (SNS) activity and diminished parasympathetic nervous system (PNS) activity, resulting in increased heart rate, blood pressure, muscle tension, changes in blood flow and perspiration, and cortisol levels (Everly and Lating 2013). However, although cortisol level appears to be the most reliable indicator, the extensive sampling process required (multiple blood or urine samples throughout the day) poses a challenge in short‐term psychological studies (Nater et al. 2013).

While the identification of physiological stress biomarkers has been largely investigated in laboratory settings using physiological or psychosocial stressors (Al’Absi et al. 1997; Cohen et al. 2000; Oswald et al. 2004), assessing psychological stress in everyday life such as work‐related pressures, social interactions, and environmental factors remains challenging (J. Weber et al. 2022). This challenge is due to the complexity of real‐world environments, individual responses, measurement inconsistency, and technological accuracy and reliability limitations (Trull and Ebner‐Priemer 2014).

Heart rate variability (HRV) reflects the fluctuation in time intervals between successive heartbeats, represented by a set of statistical metrics derived from an R‐R signal (the time interval between consecutive R‐wave peaks on an electrocardiogram) or inter‐beat interval (IBI). These fluctuations are influenced by heart‐brain interactions and autonomic nervous system (ANS) dynamics, particularly by the vagal nerve (Kim et al. 2018) (Shaffer and Ginsberg 2017). HRV can be analysed in the time‐domain, in the frequency‐domain and through non‐linear indices (Shaffer and Ginsberg 2017). In the present study, we focused on broad HRV domains, including all the above. It represents the ability of the heart to respond to physiological or environmental stimuli (Rajendra Acharya et al. 2006). Moreover, HRV has emerged as a reliable biomarker for assessing ANS activity in a non‐invasive and relatively simple manner (Kim et al. 2018). Heart rate is regulated simultaneously by the two branches of the ANS; the parasympathetic nervous system (PNS) and the sympathetic nervous system (SNS) (Shaffer and Ginsberg 2017). Changes in the activity levels of these branches are reflected through HRV and are associated with physiological stress responses (Giannakakis et al. 2019) (Thayer et al. 2012). Since the PNS regulates the SNS, HRV is considered a reliable biomarker of ANS regulatory activity in response to physiological stress (Kim et al. 2018).

To measure HRV, a high time resolution must be recorded, achieved through the gold‐standard electrocardiogram (ECG) recording or photoplethysmography (PPG) (Laborde et al. 2017).

The heart responds much faster to parasympathetic stimulation of the sinus node, which is the heart's natural pacemaker responsible for generating and regulating its rhythmic contractions, compared to sympathetic stimulation. This dominance of the parasympathetic system causes higher frequency changes or higher variability in the heart rate signal. Thus, higher HRV is typically associated with parasympathetic activation, indicating a state of relaxation—‘rest and digest’ state, while lower HRV is linked to sympathetic dominance during ‘fight or flight’ responses to a physical stress, psychological stress or challenges and during physical exercise (Rajendra Acharya et al. 2006) (Metelka 2014) (Shaffer et al. 2014).

Studies of HRV reactivity to psychological stressors in healthy human participants show that stress‐induced changes in HRV metrics correlated with a reduction in vagal tone that was quantified by decreased levels of RMSSD, pNN50, HF, increased levels of SD2/SD1 and LF/HF, and an increase in HR (Kim et al. 2018) (Laborde et al. 2017) (Delaney and Brodie 2000) (Kleiger et al. 2005). Neuroimaging studies suggest a link between HRV, and brain regions associated with reduced threat perception. HRV is seen as a mean to gauge the functional integration of brain regions involved in stress regulation and the flexible control over the ANS (Kim et al. 2018) (Thayer et al. 2012). This evidence further supports the utility of HRV as a reliable indicator of stress (C. S. Weber et al. 2010).

Although HRV is a well‐established measure of physiological stress, it lacks strong generalisability from group‐level findings to individual‐level assessments (Fisher et al. 2018) (Hoemann et al. 2023). Variables such as age (Nunan et al. 2010), sex (Koenig and Thayer 2016), and health status (De Meersman 1993) are crucial for interpreting HRV measurements, particularly for ultra‐short‐term (UST) recordings of less than 5 min (Shaffer and Ginsberg 2017), emphasising the necessity of an individualised approach in assessing stress through HRV.

HRV has several limitations that must be acknowledged. It is a complex, multifactorial signal influenced by numerous physiological and behavioural variables beyond stress, such as physical activity, respiration, hydration, and emotional arousal—whether negative or positive (Fatisson et al. 2016; Quintana and Heathers 2014; Laborde et al. 2018). Moreover, increasing evidence suggests that the low‐frequency (LF) component and the LF/HF ratio should not be interpreted as direct indices of sympathetic nervous system activity or ‘sympatho‐vagal balance’ (Billman 2013) (Heathers 2012). These findings highlight the importance of interpreting HRV within a well‐defined theoretical framework and designing studies that minimise confounding factors influencing HRV as much as possible.

Thus, we maintain that HRV can be a valid, methodologically feasible, and informative proxy of a short‐term physiological marker of stress, when used cautiously and within a well‐constructed design. In our study design, we took deliberate steps to minimise potential confounding factors by controlling physical activity, food intake, and hydration. The experimental tasks were structured to assess changes in HRV parameters over a short timespan. Given the controlled setting and limited duration, we interpret these changes as primarily reflecting stress‐related physiological responses rather than longer‐term influences.

A considerable number of studies have been conducted in the last few years to validate heart‐rate direct measurements of commercially available wrist‐worn devices from manufacturers such as Apple, Fitbit, Garmin, Polar, and Samsung (Evenson and Spade 2020; Hajj‐Boutros et al. 2022; Shcherbina et al. 2017) and presented a wide range of accuracies. Furthermore, recent research evaluated the validity of HRV direct measurements provided by wrist‐worn PPG technology compared to measurements extracted from ECG signals (Hernando et al. 2018; Nuuttila et al. 2021; Umair et al. 2021). Regardless, many of these wearables provide composite health scores (CHS)—inferred metrics that rely primarily on HRV and heart rate, but are generated through unpublished algorithms to calculate physiological derivatives (e.g., stress, body battery), the validity of which remains questionable (Doherty et al. 2025) (Peake et al. 2018).

The Garmin Vivosmart 4 fitness tracker (GV4) (Garmin Ltd., Olathe, Kansas) is a wrist‐worn device equipped with a PPG sensor designed for monitoring physiological indicators such as heart rate, VO2 max, as well as monitoring sleep, stress levels, and activity tracking. It generates a stress level score ranging from 0 to 100 every three minutes, utilising a non‐invasive method largely based on HR and HRV indices (J. Kettunen Saynatsalo (FI); and S. Saalasti Jyvaskyla (FI) 2004). Despite these features, the device does not provide access to raw IBI data or detailed HRV metrics.

Although the GV4 is widely used in both experimental research and real‐world settings, the validity of its measurements remains under scientific scrutiny (Lu et al. 2025) (Schoenmakers et al. 2025). The challenge is compounded by the restricted access to raw data, including IBI signals, which are necessary for thoroughly assessing the accuracy of these scores. In fact, certain studies have indicated that while HRV measurements under rest conditions demonstrate satisfactory accuracy, their reliability diminishes in dynamic scenarios, such as during cognitive and emotional stress (Menghini et al. 2019).

Therefore, the objectives of the current study were (1) to predict psychological stress from physiological measurements, and (2) to assess the stress level scores given by the wrist‐worn fitness tracker GV4 (Garmin Ltd., Olathe, Kansas) ability in identifying a physiological response to mental stress. Our study was designed to validate the Garmin Stress Score (GSS) for use as an experimental measure in future research, employing a validation process modelled after the benchmarking framework proposed by (Kleckner et al. 2021) (van Lier et al. 2020).

We aimed to compare and evaluate Garmin's stress score data against HR and HRV metrics as references extracted from IBI data derived from electrocardiogram (ECG) recording, using a validated heart rate monitor—Polar H10 chest strap (Polar Electro Oy, Kempele, Finland). Next, the physiological measurements were used to predict mental self‐reported stress, enabling us to monitoring real‐time perceived stress. We chose the GV4 for its lightweight design and minimal interference with participants' daily lives, including during sleep. Additionally, it maintains participant privacy by not independently measuring or storing location data, as it requires a smartphone GPS for location tracking. The study was pre‐registered on the Open Science Framework platform (Foster and Deardorff 2017) based on a pilot sample and a power analysis.

In our experiment, participants wore a Garmin Vivosmart 4 fitness tracker and a Polar H10 chest strap heart rate monitor simultaneously to record heart rate signals during a laboratory session. During that time, participants underwent tasks that were designed to elicit varying levels of mental stress. HRV metrics were calculated from the heart rate signal recorded using the Polar H10 and stress level scores provided by the Garmin Vivosmart 4 fitness tracker. Changes in GSS across stress/rest conditions were compared with the changes in HR and HRV metrics to assess validity. We hypothesised (and pre‐registered) that Garmin's stress level score would identify stress and relaxation states based on HRV‐derived calculations and will reflect higher values during the mental stress task with respect to baseline and recovery conditions. Moreover, it would correlate with heart rate (HR) and heart rate variability (HRV) measures.

2. Methods

2.1. Participants

We collected data from 60 valid participants (40/20 females/males, mean age = 27.5 ± 5.6, range 19–39). Participants were University student volunteers. All participants were healthy, had normal or corrected‐to‐normal vision, normal hearing, reported no neurological conditions, did not consume any psychiatric medications or drugs, and were not diagnosed with cardiovascular irregularities. Female participants declared that they were not pregnant. Ethical approval was obtained from the local University ethics committee, and all participants provided informed consent. Participants were compensated with a payment of 40 NIS per hour with a bonus payment ranging from 0 to 20 NIS was awarded according to individual success in the stress task.

The required sample size for the main experiment was driven from a pilot study (N = 29) conducted in the lab prior to the main experiment data collection (see ‘pilot and pre‐registration’ section). A total of 90 individuals volunteered to participate in the study. 30 participants were excluded from further analyses due to the following preregistered reasons, yielding a final valid sample size of 60:

• During the procedural phase, seven volunteers were unable to undergo the second session due to personal reasons.

• Following data collection and initial processing, 23 participants were excluded from the analysis due to missing data points. Participant exclusion primarily stemmed from technical issues encountered during data collection, such as Bluetooth disconnections between the Polar chest strap and the smartphone application, resulting in an incomplete signal necessary for the calculation of HRV metrics. Additionally, a subset of participants exhibited too many missing data points in the GSS data. Participants were deemed ineligible for analysis if there were no recorded stress points during the entirety of one of the tasks. We included only participants with a sufficient number of data points, as the experiment records only 15 points in total per participant, and excessive missing data would prevent further analysis in subsequent analyses.

• One participant exhibited a high resting HR (> 100 bpm), but their data were retained as they did not meet our predefined exclusion criteria. Notably, in a post‐hoc analysis excluding this participant, did not alter the results.

Participant's demographic information and other lifestyle characteristics are provided in Table 1.

TABLE 1.

Participant's demographic information and lifestyle characteristics—main experiment.

Numeric variables
	Mean	Std	Min	Max
Age [years]	27.5	5.6	19	39
Height [cm]	167.9	9.8	150	190
Weight [kg]	68.0	15.9	45	110
Typical sleep duration [hours]	7.1	1.3	5.0	9.0
PSS‐14 score	26.2	8.8	11	46

Categorical variables
	Count	Percentage
Gender
Female	40	66.66%
Male	20	33.34%
Smoking
Yes	3	5.00%
No	57	95.00%
Oral contraceptive intake
Yes	33	55.00% (82.50% of females)
No	27	45.00% (17.50% of females)
Exercise regularly	37	61.67%
Exercise frequency (times per week)
Low (1–2)	17	28.33%
Medium (3–4)	15	25.00%
High (5 or more)	5	8.33%
Exercise intensity
Low	6	10.00%
Medium	21	35.00%
High	10	16.67%

2.2. Procedure

The procedure included a preparatory meeting and an experimental task session. During the preparatory meeting, participants were provided with a GV4 fitness tracker synchronised to a designated personal account in the Garmin Connect mobile application on their phones. Demographic information including age, sex, height, and weight was collected. Additional behavioural information including smoking habits, oral contraceptive use, habitual exercise (frequency and intensity), and habitual sleep routines were collected (see Table 1 for details). This information was used to explore how personal differences influence the physiological metrics assessed. Inclusion criteria were verified and informed consent was obtained after explaining the experimental procedures. To ensure accurate identification of stressful moments, participants were instructed to wear the fitness tracker for one full day and night prior to the main laboratory session, as per the recommendations outlined in the GV4 manual. Participants were instructed to abstain from food, caffeine, and smoking for at least 2 hours before the experimental task session, as well as intense physical activity, drugs, and alcohol in the 24 h prior to the experimental task session while maintaining their regular sleep routine.

Participants came back to the laboratory for the experimental task session the following day or later in the week. At the beginning of the session, the experimenter ensured that stress and heart rate data had been collected using the GV4 during the preceding day and night. This laboratory‐based session comprised three consecutive tasks each lasting 15 min (vs. 30 min in the pilot experiment): a resting phase (baseline), a stressful phase, and a relaxation phase (recovery), for a total duration of 45 min.

Throughout the session, participants remained seated in front of a computer in a quiet laboratory testing room, wearing both the GV4 fitness tracker and the Polar H10 chest strap. They were instructed to minimise movements during the tasks, while data were simultaneously collected from both devices. During the resting phase, participants were instructed to sit quietly without any external stimuli. In the stressful phase, participants engaged in a computerised mental arithmetic task modelled after the ‘Montreal Imaging Stress Task’ (MIST), a protocol used to induce psychosocial stress in participants (Dedovic et al. 2005). Before they began, participants were informed of a monetary incentive contingent upon their performance. During the relaxation phase, participants viewed a video featuring natural landscapes and listened to calming music. Furthermore, participants were asked to provide self‐reported stress assessments. Participants verbally reported their subjective assessment of mental stress levels on a scale from 1 to 10 after completing each task. This information provided additional insight into participants' real‐time stress experiences throughout the three mental states and enabled us to further investigate, as well as predict, self‐reported mental stress from physiological measurements.

In addition, following the completion of the three phases, participants completed a Perceived Stress Scale questionnaire (PSS‐14) (Cohen et al. 1983). The PSS‐14 focuses on how uncontrollable or overwhelming respondents found their lives in the recent past, typically within a 1‐month period. It contains 14 items that ask about thoughts and feelings experienced recently, providing a broader view of overall stress. A visual representation of the experimental protocol is presented in Figure 1.

FIGURE 1.

Protocol flowchart of the main experiment.

2.3. Data Collection and Wearables

To explore the relationship between physiological and psychological stress level scores, data were collected concurrently using the GV4 and a Polar H10 chest‐strap heart rate monitor (Polar Electro Oy) during the main laboratory session. The Polar H10 served as a validated device for recording R‐R signals extracted from the QRS complex using electrocardiography (ECG) technology. Stress level scores came from the GV4 which uses a PPG sensor.

The Garmin Connect mobile application (Garmin Ltd., Olathe, Kansas) was utilised to sync and store the stress data recorded by the GV4. Since direct access to the detailed data from the Garmin Connect app is restricted, we used Fitabase (Small Steps Labs LLC). The R‐R signal was recorded using a Polar chest strap, with data captured and retrieved through Elite HRV installed on an iPhone13 Pro, a mobile application designed for HRV monitoring. The data was then used for further analyses, calculating HRV metrics using Kubios HRV Premium software (version 3.5.0; Biosignal Analysis and Medical Imaging Group, Department of Physics, University of Kuopio, Kuopio, Finland) (Tarvainen et al. 2014).

Reducing the duration of each task in the main experiment from 30 to 15 min relative to the task duration in the pilot procedure was implemented to minimise participant fatigue and discomfort, particularly during the stress‐inducing task. Prolonged exposure to the stressor could lead to reduced engagement and potentially confound the physiological responses being measured. By shortening the task duration, we aimed to maintain participant compliance and data quality while still eliciting measurable stress responses.

For more details, please see supporting information, section C.

2.4. Pilot Experiment and Pre‐Registration

A pilot experiment was conducted to assess the experimental protocol's feasibility and estimate the main experiment's sample size. Data from 29 participants was collected and analysed. A power analysis was conducted to estimate the sample size required to detect significant differences in GSS, HR, and RMSSD values between recovery and stress conditions. The analysis yielded a sample size of 60 participants (or below) for the main experiment with a statistical power of 0.80 and a significance level of 0.0125 (applying Bonferroni correction to account for multiple comparisons of HRV metrics p < 0.05/4 = 0.0125). Further details are available at the ‘Sample size rationale’ section in the pre‐registration on the OSF platform. Following the analysis of the pilot data, minor modifications were made to the experimental tasks and data collection to enhance variable control and improve data quality in the main experiment. Task durations were reduced to minimise participant fatigue and discomfort. Participants also provided self‐reported stress assessments after each task, and additional data on sleep, fitness, and lifestyle were collected (see details in the main experiment procedure description). The main study was preregistered in OSF, outlining hypotheses, study design, and analysis plan prior to data collection, based on the pilot sample. Results of the pilot experiment are available in Section B of the supporting information. Pre‐registration is available at the OSF repository, combined with the experimental data, task codes and analyses codes: https://osf.io/gdr2n/?view_only=56e6c638a6d340c5a772670c309c0504.

2.5. Statistical Analyses

2.5.1. Confirmatory Analyses

The first part of the statistical analysis of the main experiment was pre‐registered prior to data collection, based on the analysis of the pilot study (details of the pilot study data analysis are available in section A in the supporting information). It includes two components: (1) psycho‐physiological interactions, and (2) device comparison.

For the psycho‐physiological interaction, we first started by examining the temporal aspect of stress reactivity. Next, we used paired t‐tests to compare HRV parameters, HR, and GSSs between stress conditions and both baseline and recovery phases, to identify stress‐related physiological responses to the stressful phase. For this test we used the mean values of each metric for each condition. We tested the normality assumption prior to conducting paired t‐tests using the Shapiro‐Wilk test. Three parameters (RMSSD, GSS, and LF/HF) did not meet the normality assumption. Therefore, we additionally applied the Wilcoxon signed‐rank test as a non‐parametric alternative. Statistical significance was determined with p values < 0.05. To account for multiple comparisons (of HRV metrics), a Bonferroni correction was applied, setting the significance level to p < 0.05/4 = 0.0125.

For the device comparison part, we started with a correlation test between all given records to examine inter‐correlation between variables. Next, we employed two methods of correlational analysis. The first method involved a correlational analysis of condition means. For this analysis, we calculated the mean GSSs and HRV metrics within each experimental condition (baseline, stress, and recovery) for each participant. Using these means, we computed the Pearson correlation coefficient for each condition based on the data from all 60 participants. This method allowed us to assess the strength and direction of the relationship between HR, HRV parameters, and GSSs across different states. The second method was a within‐subject analysis. For each participant, we calculated the Pearson correlation coefficient between GSSs and each HRV metric. We then averaged these individual correlation coefficients across all participants to obtain a mean correlation coefficient for each HRV metric. To determine the significance of these correlations, we conducted a one‐sample t‐test on the mean correlation coefficients. This process resulted in a table displaying the mean Pearson's r and associated p‐values for each HRV metric with GSSs, indicating whether there was a meaningful relationship between GSSs and HRV metrics across the participant sample. While more sophisticated analytic models could offer additional insights, this two‐pronged approach provided a multilayer, comprehensive evaluation of the association between GSSs and HRV metrics, both at the group level and within individual participants.

2.5.2. Exploratory Analysis

Following the pre‐registered analysis, we employed two additional analyses that were not pre‐registered: (1) examining the predictive power of physiological measurements on psychological (self‐reported) stress, and (2) the effect of individual differences (e.g., sex, age, exercise habits) on HR, HRV, and GSS in stress reactivity.

To assess the predictive power of physiological measurements on self‐reported stress, we applied a linear mixed‐effects model, with reported stress as the dependent variable, the physiological parameters (heart rate, RMSSD, SD2/SD1 ratio, LF/HF ratio, HF power, and GSS) as fixed effects, and participant as a random effect. To address collinearity among predictors, we additionally performed stepwise regression (both forward and backward selection (Sanchez‐Pinto et al. 2018),) using the same variables.

For the individual differences we used linear mixed‐effects models (LMM) to assess the impact of physiological and demographic factors on GSS across different experimental conditions (baseline, stress, and recovery). The models incorporated independent variables such as sex, age, body mass index (BMI) in [kg/m²], regular exercise habits (at least 1–2 times a week or none), average nightly sleep duration in hours, oral contraceptive use, and baseline physiology categories (low or high RMSSD, HR and GSS). Two groups of the physiological category were defined based on a median split on resting values (high vs. low RMSSD, HR and GSS, taken during baseline condition). Additionally, interaction effects between sex and stress condition, and between physiology category and stress condition were investigated to test if the effect of the stress task differs based on sex or if individuals with low versus high at baseline respond differently to the stress exposure. This type of model allowed us to capture both the fixed effects of the experimental conditions and other covariates, as well as the random effects due to individual variability. A model was built for each physiological metric of interest as the dependent variable: mean HR (bpm), RMSSD (ms), SD2/SD1, HF power (nu), and LF/HF. The analysis was conducted using the mixed linear model regression approach utilising the ‘statsmodels’ package on Python. Significance was determined at p < 0.05, and coefficients with their respective p‐values were examined to ascertain the strength and direction of the associations.

3. Results

3.1. Confirmatory Analyses

3.1.1. Variation of Measures Across Experimental Conditions: Psycho‐Physiological Interaction

We began by visualising the mean data over time for each physiological measure to examine potential trends in relation to psychological stress. We found that GSS and heart rate responded immediately to changes in reported stress, whereas no measure exhibited a delayed trend (see Figure 2).

FIGURE 2.

Mean trajectories of physiological parameters across time. Each cell showed one physiological parameter. Bold lines represent the mean across the group, with shaded areas representing the standard error. Vertical dashed lines indicate task boundaries (baseline/rest, MIST, and recovery, respectively).

Then, paired t‐tests assessed significant differences in participants' physiological measures across the three conditions (Baseline, Stress, and Recovery). A significance threshold was set at p < 0.0125 after applying the Bonferroni correction to account for multiple comparisons of HRV metrics.

The results of this analysis revealed significant differences in Mean HR between Stress versus Baseline (p < 0.001) and Stress versus Recovery (p < 0.001), indicating elevated HR during stress‐inducing tasks compared to baseline and recovery periods. No significant differences were observed in the comparison of HR during Recovery versus Baseline. Mean RMSSD decreased during the stress condition relative to baseline and increased during recovery, however, this difference did not meet the significant threshold. The SD2/SD1 Ratio exhibits significant difference between Stress and Baseline (p = 0.0014). HF Power (nu) decreased significantly during the stress task compared to baseline (p = 0.003). LF/HF increased during the stress condition relative to baseline, but this difference did not meet the significant threshold. The GSS showed a very similar pattern to HR with a significant increase during the stress task (p < 0.001) and a significant decrease during recovery (p < 0.001). These findings underscore the dynamic physiological responses to stress‐inducing tasks and subsequent recovery periods. See Figure 3A–F for a visual presentation of the comparison. For full results, please refer to section C of the supporting information: Table S IV.

FIGURE 3.

The distribution of participants' HR (A), GSS (B), and HRV parameters (C–F) and self‐reported stress levels (G) across three conditions: baseline (rest), stress, and recovery. Notably, mean values in each experimental condition are indicated within the boxplot. Significant differences (p value < 0.0125) across conditions are marked by an asterisk.

Alongside physiological measures, a paired t‐test was conducted to examine differences in subjective self‐reported stress levels (see Figure 3G). On average, participants reported higher levels of subjective stress following the stress task compared to both baseline and recovery conditions (p < 0.001).

3.1.2. Correlation Analysis: Device Comparison

A correlation analysis, considering all available data for each metric, revealed strong associations between certain physiological parameters. GSS was highly correlated with HR (r = 0.84), and the LF/HF ratio correlated strongly with the SD2/SD1 ratio (r = 0.78). Overall, physiological measures were moderately to strongly interrelated (mean absolute r = 0.597 ± 0.186). In contrast, self‐reported stress showed consistently weak correlations with both physiological and CHS (mean absolute r = 0.137 ± 0.053), indicating a limited correspondence between subjective and physiological indices of stress (see Figure 4).

FIGURE 4.

Correlation heatmap of physiological parameters and self‐reported stress. Colour intensity reflects the correlation coefficient (r), with red indicating positive and blue indicating negative correlations.

Next, we employed two different methods of computation, found consistent associations between the GSSs and various physiological metrics: within‐condition and within‐subject correlations.

3.1.2.1. Within‐Condition Correlation

Pearson correlation coefficients were computed from mean scores across participants, within each condition (each participant had one observation in each condition, totaling three observations for participant). Robust positive correlations emerged between GSSs and Mean HR, as well as SD2/SD1, with coefficients ranging from 0.84 to 0.85 (p < 0.0001) and 0.61 to 0.64 (p < 0.0001), respectively. Conversely, a notable negative correlation was observed between RMSSD and Garmin's metric, with coefficients ranging from −0.59 to −0.63 (p < 0.0001). Furthermore, lower positive correlations were found between GSSs and LF/HF (0.32 (p = 0.0126) to 0.38 (p = 0.0028)) and moderate negative correlations with HF (−0.4 (p = 0.0014) to −0.43 (p = 0.0006)). Comprehensive analysis data, including Pearson correlation coefficients and corresponding p‐values for each metric across all conditions, can be found in section C of the supporting information: Table S V.

3.1.2.2. Within‐Subject Correlation

We investigated the relationships between GSSs and the physiological parameters at the individual level. Individual correlations per participant were computed, and the averages across participants were analysed. This approach yielded a significant positive mean correlation between GSS and Mean HR (r = 0.74, SD = 0.35, p < 0.0001), indicating a consistent association between higher stress level scores provided by Garmin and elevated HRs. RMSSD displayed a lower negative correlation (r = −0.41, SD = 0.42, p < 0.0001), suggesting that increased GSSs were associated with reduced HRV. SD2/SD1 demonstrated a lower positive correlation (r = 0.32, SD = 0.45, p < 0.0001), indicating a tendency for higher GSSs to coincide with higher SD2/SD1 ratios. LF/HF and HF power exhibited low correlations with Pearson's r of 0.18 (SD = 0.41, p = 0.001) and −0.20 (SD = 0.42, p = 0.001), suggesting weaker associations between Garmin's metric and these metrics. High standard deviations were observed in the correlation statistics, indicating variability in these measures among participants. These findings highlight the complex interplay between physiological responses and stress levels, as captured by GSSs. A summary table of the results can be found in section C of the supporting information: Table S VI.

3.1.3. Correlation Between Measurements and General Indices

We examined correlations between the PSS‐14 questionnaire score (a widely used questionnaire that assesses individuals' perceived stress levels over the past month), and the physiological and self‐reported measures.

Results revealed no significant relationships between the measures and PSS‐14 scores. Pearson's correlations were weak and non‐significant: HR (r = −0.03, p = 0.797), RMSSD (r = −0.04, p = 0.769), SD2/SD1 (r = −0.06, p = 0.632), LF/HF (r = −0.07, p = 0.624), HF power [nu] (r = 0.01, p = 0.919), GSS (r = −0.17, p = 0.201), and self‐reported stress (r = 0.25, p = 0.055). Similarly, the stress to baseline differences showed weak and non‐significant correlations: HR (r = −0.15, p = 0.266), RMSSD (r = 0.16, p = 0.229), SD2/SD1 (r = −0.11, p = 0.414), LF/HF (r = −0.13, p = 0.321), HF power [nu] (r = 0.04, p = 0.791), GSS (r = −0.19, p = 0.146), and self‐reported stress (r = 0.11, p = 0.394). These findings suggest no meaningful relationship between the examined metrics and perceived stress levels as measured by the PSS‐14 in this dataset. We further investigated the relationship between participants' performance in the arithmetic task conducted during the stressful phase and their physiological metrics, as well as the disparity in physiological measures between conditions.

The analysis revealed no significant relationships. Pearson's correlations for baseline measures were weak: HR (r = −0.06, p = 0.670), RMSSD (r = 0.05, p = 0.720), SD2/SD1 (r = −0.00, p = 0.984), LF/HF (r = 0.09, p = 0.475), HF power [nu] (r = 0.04, p = 0.764), Garmin stress (r = −0.05, p = 0.684), and reported stress (r = −0.19, p = 0.140). Similarly, the correlations for the stress to baseline differences were also non‐significant: HR (r = −0.03, p = 0.829), RMSSD (r = 0.00, p = 0.979), SD2/SD1 (r = −0.06, p = 0.637), LF/HF (r = 0.06, p = 0.644), HF power [nu] (r = −0.04, p = 0.792), Garmin stress (r = 0.20, p = 0.131), and reported stress (r = −0.09, p = 0.486). These findings suggest no meaningful correlation between the performance in the arithmetic task and the physiological measures or the change in it across stress/rest state.

3.2. Exploratory Analysis

3.2.1. Predicting Reported Stress From Physiological Measurements

3.2.1.1. Physiological Predictors of Self‐Reported Stress

We used a linear mixed‐effects model to estimate the predictability of subjective reported stress from physiological parameters. Heart rate emerged as the only parameter significantly associated with reported stress (β = 0.067 ± 0.032, p = 0.039), while GSS showed a marginal effect (β = −0.023 ± 0.012, p = 0.053). All HRV measurements were found to be not significant, with mixed directions of coefficients. See Figure 5 for results visualisation.

FIGURE 5.

Forest plot showing linear mixed‐effects model results for predictors of self‐reported stress. Variables are ordered by statistical significance. Point estimates (dots) with 95% confidence intervals are shown. p‐value < 0.1 marked in +, p‐value < 0.05 marked in *.

Given the high collinearity among predictors, we applied forward and backward feature selection. Only heart rate was retained, yielding (β = 0.038, p = 0.005), consistent with the results from the full model.

3.2.1.2. Individual Differences in Self‐Reported Stress

To examine how individual characteristics affect self‐reported stress, we applied a linear mixed effect model. The model included four main factors: sex (male vs. female), baseline HRV category, baseline heart rate category and baseline GSS category. HRV, HR and GSS categories were a binary classification based on median value of the parameter during the baseline phase. All factors were allowed to interact with the experimental conditions to explore their combined influence on the dependent variable. To adjust for potential confounders, age, BMI, oral contraceptive use, and sleep duration were included as covariates. Additionally, a random intercept was incorporated for each participant to account for within‐subject variability across repeated measures.

A significant effect of sex was identified, with males self‐reporting significantly lower stress levels compared to females (β −1.200, p = 0.028), see Figure 6A. Other individual characteristics were found to be non‐significant. While examining the effect of physiological categories (high and low baseline value of HRV, HR and GSS), only HR showed significant effect (β −0.70, p = 0.03), meaning that participants with lower HR atin baseline reported less stress (see Figure 6C). No such effects were detected for HRV and GSS. Results are shown in Figure 6.

FIGURE 6.

Self‐Reported Stress Across Experimental Conditions, divided by Sex (Female vs. Male), and physiological categories (high and low baseline classification)—HRV, heart rate and Garmin stress score (GSS).

3.2.2. The Role of Individual Differences in Moderating Physiological Variables

We examined whether certain individual differences moderate the physiological measures we assessed. We tested the effect of sex, age, BMI, habitual exercise, average nightly sleep duration, oral contraceptive use, and baseline HRV (low vs. high based on a median split of resting RMSSD) on GSS, HR, HRV, and self‐reported stress across the three conditions (baseline, stress, and recovery) through linear mixed‐effects models (LMEM). We used the same model’ predictors as in Section 3.2.1.2. Separate linear mixed‐effects models were developed for each dependent variable: mean HR, and HRV metrics.

The split into two groups, low versus high baseline HRV, was performed by applying a median split of resting RMSSD values (median = 35.75 ms, during baseline condition). These groups, representing 30 participants each, differed significantly in RMSSD (Low HRV group: 22.2 ± 6.6 ms; High HRV group: 60.2 ± 23.6 ms, t (Park et al. 2014) = 8.4, p < 0.001).

3.2.2.1. Stress Task Effect

The LMEM model revealed significant differences in the physiological metrics between conditions. Specifically, higher HR and SD2/SD1 values were observed during the stress task compared to baseline (β = 6.305, p < 0.001 and β = 0.370, p < 0.001, respectively), while RMSSD was lower (β = −7.943, p = 0.001). In the frequency domain, HF levels decreased during the stress task compared to baseline (β = −8.505, p < 0.001) and, although they increased during recovery, they remained lower than baseline (β = −5.028, p = 0.023). The LF/HF ratio increase during the stress task compared to baseline was statistically insignificant. A significant increase in GSS was observed during the stress task (β = 10.384, p < 0.001), with a marginally higher GSS also noted during recovery (β = 3.908, p = 0.048), both compared to baseline. Finally, participants reported higher stress levels at the end of the stress condition compared to baseline (β = 2.764, p < 0.001), emphasising the stress‐induced task efficacy.

3.2.2.2. Garmin Stress Score—GSS

The LMM results revealed several significant predictors of GSS, including HRV category, exercise, BMI, and sleep duration. Participants with low baseline HRV exhibited significantly higher GSS compared to those with high HRV (β = 21.679, p < 0.001). Regular exercise and longer sleep durations were associated with lower stress scores captured by the GSS (β = −17.948, p = 0.002 and β = −4.304, p = 0.027 respectively). Higher BMI was associated with higher stress scores (β = 1.365, p = 0.018). No significant effects related to sex, oral contraceptive use, or age were found in Garmin stress scores. Interactions between sex and experimental conditions, as well as HRV category and experimental conditions, are not significant. Figure 7 demonstrates GSS across experimental conditions, categorised by individual differences such as Sex, Habitual Exercise, and Baseline HRV Category.

FIGURE 7.

GSS Across Experimental Conditions, divided by individual differences including (A) sex (Female vs. Male), (B) exercise regularly (‘True’ for at least 1–2 times a week vs. ‘False’ for none), and (C) HRV category (low vs. high baseline RMSSD).

3.2.2.3. Heart Rate Response

The factors identified to influence HR significantly include baseline HRV category and habitual exercise. Specifically, participants categorised with low HRV exhibited higher HR compared to those with high HRV (β = 12.057, p < 0.001). Additionally, individuals who exercise regularly (n = 37) have, on average, a lower HR (β = −9.830, p = 0.001) compared to those who do not exercise (n = 23). The within‐subject random effect was significant (β = 0.158, p = 0.004), indicating variability in HR across participants. Results are shown in Figure 8.

FIGURE 8.

Mean HR Across Experimental Conditions, Divided by Individual Differences including sex (Female vs. Male), exercise regularly (‘True’ for at least 1–2 times a week vs. ‘False’ for none regularly), and HRV category (low vs. high baseline RMSSD).

3.2.2.4. HRV Metrics

The mixed linear model analysis revealed that baseline HRV category influenced all the HRV variables. Sex influenced the frequency domain metrics. The interaction effect of sex with experimental condition moderates SD2/SD1 ratio and HFnu. As expected, the analysis indicates that participants categorised with low baseline HRV demonstrated a significantly lower RMSSD (β = −36.442, p < 0.001) and lower levels of HF (nu) (β = −17.621, p < 0.001). while low baseline HRV was significantly correlated with higher levels of SD2/SD1 ratio (β = 0.843, p < 0.001) and increased LF/HF ratio (β = 2.076, p < 0.001). See Figure 9 for a visual presentation of the differences.

FIGURE 9.

HRV metrics Across Experimental Conditions, divided by HRV category (low vs. high baseline RMSSD).

A significant interaction effect of the HRV category with the stress condition in moderating RMSSD was identified. The two groups—low versus high baseline HRV display an opposite trend of RMSSD values across conditions (as shown in Figure 9A). Participants with high baseline HRV (n = 30) displayed a decrease in RMSSD (approximately 8.2 points) during the stress task, followed by a slight increase during recovery. While participants with low baseline HRV (n = 30) exhibited a marginal increase during the stress task (approximately 1.1 points). Yet, the within‐group differences did not reach statistical significance.

An interaction effect of sex and the stress condition on SD2/SD1 ratio was significant with males showing a reduction in the SD2/SD1 ratio under stress compared to females who on average show increased values during the stress task (β = −0.344, p = 0.003). Sex had a significant effect on the frequency domain metrics: LF/HF ratio and HFnu. Males had higher levels of LF/HF (β = 1.772, p = 0.008), and lower levels of HFnu (β = −14.586, p = 0.001). A significant interaction between sex and condition suggests that the relative change in HFnu from rest to stress task is influenced by sex, with males experiencing an increase in HFnu during stress condition compared to females who experiencing a decrease (β = 8.661, p = 0.003). See Figure 10 for a visual representation.

FIGURE 10.

HRV metrics Across Experimental Conditions, divided by Sex (Female vs. Male).

4. Discussion

In the present study we measure perceived psychological stress and collected physiological data from wearable devices (Garmin and Polar H10), aimed to predict stress levels from physiological measurements, as well as assess the stress level scores. Garmin Vivosmart 4 fitness tracker provided a stress score (GSS, a customer health score), and physiological data was recorded by the Polar H10 chest strap. Psychological stress was manipulated during varying stress/rest conditions, as participants were performed in three lab phases: baseline (rest), stress (arithmetic task), and recovery (relaxing video/music). Each of the experimental conditions lasted 15 min during the main experiment, and 30 min during an earlier pilot study.

We acknowledge that the body's response to physiological stress is inherently complex and multifaceted. However, to maintain methodological feasibility, we limited the number of physiological measures collected. HRV parameters (measured by a chest‐strap sensor), alongside heart rate (measured by Garmin Vivo 4 smartwatch) were selected as our primary indicators due to their accessibility, non‐invasiveness, and established relevance to physiological stress, despite the limitations.

Our pre‐registered analysis (based on the pilot study) revealed a dynamic response of physiological parameters to the stress‐inducing task, as evidenced by significant differences in mean HR, SD2/SD1 ratio, and HF power (nu) between stress conditions and baseline/rest periods. In addition, subjective self‐reported stress levels were higher at the end of the stress‐inducing task, supporting protocol validity. These results align with previous studies indicating alterations in autonomic regulation during mental stress (Kim et al. 2018), specifically increased sympathetic activity and decreased parasympathetic activity. GSSs were significantly higher during the stress condition relative to baseline and recovery phases, demonstrating that it successfully reflects the psychological stress response. Additionally, our data revealed associations between GSS and the physiological parameters, specifically a high correlation with HR and moderate correlations with SD2/SD1 and RMSSD. These results shed light on the potential utility of the wearable device in assessing stress levels remotely in natural settings.

In an exploratory analysis, we found that heart rate demonstrated the most consistent association with real‐time self‐reported stress, as verbally indicated by participants during the experiment.

The underlying motivation for this study was to understand what is the meaning of the widely used ‘Stress’ value in the most common wearable used. Despite the notion that PPG technology has been validated, and Garmin is the most sold device, the stress score has never been validated, and it seems impossible to re‐create it from the patents. Since the release of our preprint, a peer‐reviewed study has been published indicating that GSS reflects arousal levels rather than stress per se (van der Mee et al. 2025). This observation aligns partially with our findings, as both studies demonstrate that the GSS reliably tracks autonomic activation associated with stress tasks but does not clearly distinguish between stress‐related and non‐stress‐related arousal. Taken together, these results suggest that while the GSS provides a robust physiological index of activation, its interpretation as a direct measure of ‘stress’ should be approached with caution and contextualised within the broader psychophysiological state of the individual.

Under the context of naturalistic validity, our validation study is a necessary step for any study performed with customer health scores (CHS), including stress scores. These bracelets only measure ‘stress’ when participants are stationary and thus there were no special laboratory conditions. Participants received the bracelets 24 h prior to the laboratory testing to establish a pre‐lab adaptation phase, and thus it was performed partly in the field, and no special instructions were given. We performed the study in the lab to be able to perform the concurrent measurement with the commercial Polar strap and iPhone that recorded ECG for HRV calculation.

The scientific potential of these devices makes them allure and the ability to use them to assess physiological biomarkers outside of the laboratory, under ecologically valid conditions is valuable. However, it is worth noting that GSS is not calculated continuously (e.g., during movement). Future studies could address this ecological condition by testing its reliability under ‘free‐living’ circumstances, or by implementing specific procedures and study designs to overcome this limitation.

4.1. CHS and Physiological Responses to Induced Mental Stress

Both the pilot and pre‐registered (main) experiment consistently observed a parallel pattern between GSS and mean HR across the baseline, stress, and recovery phases, demonstrating a significant increase during the stress task and a decrease during recovery. Furthermore, among HRV metrics that were assessed (RMSSD, SD2/SD1, LF/HF, HF power), SD2/SD1 exhibited a significant increase during the stress task in both the pilot and main experiments. However, changes in RMSSD were not statistically significant. Among the spectral analysis metrics, results from the main experiment demonstrated a significant decrease in HF Power during the stress condition, while changes in LF/HF ratios were not statistically significant in either experiment. Furthermore, self‐reported subjective stress levels reported by participants after the stress‐inducing task were significantly higher compared to levels reported right after both baseline and recovery phases, supporting the efficacy of the arithmetic task as an effective tool for eliciting perceived stress among participants.

The significant increase in mean HR during the stress task compared to baseline and recovery is consistent with sympathetic nervous system activation commonly associated with heightened physiological arousal or stress responses (Kim et al. 2018) (Taelman et al. 2009) (Fechir et al. 2008). The decrease in HF power during the stress condition is consistent with parasympathetic nervous system withdrawal associated with stress responses (Kim et al. 2018) (Laborde et al. 2017) (Melillo et al. 2011) (Porges 1995).

An increase in SD2, associated with long‐term variability, may indicate sympathetic activation, while decreases in SD1, linked to short‐term variability, may reflect reduced parasympathetic activity (Roy and Ghatak 2013). Therefore, an elevated SD2/SD1 ratio could suggest a shift in autonomic regulation toward sympathetic dominance.

The lack of significant difference in RMSSD and LF/HF across conditions in the univariate analysis, in both the pilot and pre‐registered experiments, warrants further investigation into the sensitivity of these metrics to the chosen stress task or to the effect of individual differences. RMSSD, as a measure of HRV, reflects the beat‐to‐beat variation in heart rate and is predominantly influenced by parasympathetic nervous system activity (Laborde et al. 2017). The lack of significant findings regarding RMSSD could be attributed to various factors including stress response complexity and individual differences. Although physiological and mental stress typically correlates with reduced parasympathetic activity and subsequent RMSSD decreases, the physiological stress response is multifaceted and may involve intricate physiological mechanisms. For instance, one study (Park et al. 2014) illustrated an interesting interplay between tonic HRV and phasic HRV and demonstrated how tonic HRV might affect the suppression or enhancement of phasic HRV. This dynamic could also extend to how individual differences in physiological stress reactivity and stress regulation (as well as coping mechanisms) might influence RMSSD levels during mentally stressful tasks, leading to variability in responses within the study sample. Additionally, temporal dynamics could offer an alternative explanation. RMSSD reflects short‐term variations in heart rate, influenced by rapid changes in autonomic activity (Malik 1996). Stress‐induced changes in autonomic regulation may have been transient or occurred at different time points during the stress task, possibly contributing to the lack of significant differences in RMSSD between conditions. Despite the absence of statistical significance, the trend of lower RMSSD during the stress condition compared to baseline aligns with theoretical expectations and previous research on stress‐induced HRV changes (Kim et al. 2018). While the lack of statistical significance may limit the interpretability of these findings, they nonetheless contribute to our understanding of the complex interplay between physiological stress and autonomic regulation.

Importantly, non‐linear HRV indices such as the SD2/SD1 ratio may, in certain contexts, provide a more sensitive and reliable reflection of the complex, non‐stationary dynamics of autonomic regulation than linear time‐domain parameters like RMSSD. This notion aligns with our findings, in which SD2/SD1 demonstrated significant variation across conditions, whereas RMSSD did not reach statistical significance. Such non‐linear parameters may better capture subtle alterations in cardiac autonomic balance during stress, reflecting the intricate interplay between sympathetic and parasympathetic influences that linear measures may underestimate.

The complexity of physiological stress responses and individual differences may contribute also to the inconsistent differences observed in LF/HF across conditions. Moreover, critics have challenged the view that the LF/HF ratio reflects a sympatho‐vagal balance due to its weak correlation with sympathetic nerve activation and the non‐linear relationship between sympathetic and parasympathetic activity (Billman 2013) (Eckberg 1997). Previous work has shown that frequency‐domain metrics, as well as the non‐linear SD2/SD1 ratio, can be influenced by complex patterns of activation between the sympathetic and parasympathetic branches of the autonomic nervous system, such as coactivation or co‐inhibition (Quigley et al. 2024). Consequently, these measures should be interpreted with caution, as their variations may reflect more than just physiological or psychological stress, potentially encompassing a broader range of autonomic processes (Smith et al. 2017). We illustrated the impact of sex on LF/HF and HF metrics. The observation that these metrics change in opposite directions in response to stressors could potentially explain the mixed results. Previous research that investigated HRV changes during mental stress has similarly reported non‐significant differences in the LF/HF ratio (Taelman et al. 2009). Therefore, although this metric is frequently utilised in HRV research, its reliability remains uncertain as well as its impact on the interpretation of our findings.

4.2. Correlations of Garmin Metric With Physiological Measurements

The correlational analysis revealed a consistently high positive correlation between the GSS and mean HR as evident in both computation methods (average of within‐subject correlations and within‐condition‐ between‐subject analysis). Results were consistent across both the pilot and the pre‐registered experiment. Conversely, correlations between the GSS and HRV metrics were comparatively lower, yet significant. In the main experiment, the Garmin metric exhibited a negative correlation with RMSSD and a positive correlation with SD2/SD1, aligning with the expected autonomic shift in response to the stress task. Conversely, associations with spectral analysis metrics (LF/HF and HF power) remained low across both experiments and computation methods, mirroring the lack of significant change observed in the variations of measures across conditions analysis.

The strong correlation of the GSS with HR might suggest that HR may carry greater weight than HRV in the Garmin algorithm's stress calculation. The high standard deviations in the correlation analyses suggest considerable variability among participants. This variability may reflect individual differences in autonomic regulation and physiological stress responses, highlighting the need for personalised approaches in HRV analysis. Furthermore, we surmise that this variability strengthens our study by capturing real‐life inter‐individual variability, reinforcing the ecological validity of our study and its applicability across diverse populations.

4.3. Predicting Self‐Reported Stress From Physiological Measurements

We also examined, in the context of ecological studies, which physiological parameter best aligned with self‐reported stress during the experiment. The parameters tested included heart rate (HR), GSS, and four HRV metrics. Among these, only HR showed a statistically significant association with reported stress, while GSS demonstrated a marginal negative effect. These results support the use of HR as a simple and well‐characterised marker for stress in ecological settings. While the Garmin Stress Score calculation algorithm relies heavily on HRV, the algorithm to detect stress/relaxation states relies more on HR (J. Kettunen Saynatsalo (FI); and S. Saalasti Jyvaskyla (FI) 2004), and thus the stress‐state may have greater potential to discriminate subjective real‐time perceived stress.

When examining the effect of individual characteristics on self‐reported stress, we found that women had higher self‐reported mental stress than men, and this is consistent with known literature (Verma et al. 2011) (Costa et al. 2021). Also, heart rate category (high or low HR at baseline) had a significant effect on self‐reported mental stress, while no other physiological parameter had shown the same effect. However, the patterns of stress formation and relaxation were different when comparing HRV and GSS. This trend can be examined more deeply in a future study.

4.4. The Role of Individual Differences in Moderating Physiological Variables

HRV and GSS measurements are influenced by numerous individual factors including demographics and lifestyle (Sammito et al. 2024). Our mixed‐effects linear model analyses revealed that sex, exercise, and tonic HRV significantly moderated HR and HRV values, while GSS was influenced by exercise, tonic HRV, BMI, and average nightly sleep duration.

Participants with higher baseline HRV displayed significantly lower HR and GSS overall, particularly during stress, consistent with existing research showing that higher tonic HRV reflects more adaptive stress responses (Kim et al. 2018) (Thayer et al. 2012) (C. S. Weber et al. 2010).

Sex differences emerged primarily in frequency domain metrics and self‐reported stress: males reported lower stress yet showed higher LF/HF ratio and lower HF values compared to females. The stress condition's impact on SD2/SD1 ratio and HFnu differed significantly between sexes, with women showing increased LF/HF and decreased HFnu during stress, while men exhibited the opposite. This pattern is also consistent with existing literature (Kim et al. 2018). These sex‐specific patterns may explain the relatively weak correlations of frequency domain metrics with GSS observed in univariate analyses.

Regular exercisers (≥ 1–2 times weekly) demonstrated lower average HR and GSS compared to non‐exercisers, aligning with established associations between fitness and autonomic function (De Meersman 1993), though exercise did not significantly influence the magnitude of stress responses.

The linear mixed model confirmed significant stress condition effects on GSS, self‐reported stress, HR, and most HRV metrics while controlling for individual characteristics. Notably, GSS remained elevated during recovery compared to baseline, indicating residual post‐stress effects, a pattern also observed in HFnu values. These findings strengthen our pre‐registered analyses and confirm both the stress task's efficacy and GSS's capability to detect stress comparably to HR and HRV measurements.

4.5. Limitations

While our findings provide valuable insights, several limitations must be acknowledged, and further exploration could provide additional evidence concerning the effectiveness of a stress score based on HRV and the GSS metrics in assessing individuals' stress responses:

• The Stress Induction Method: We relied on an arithmetic quiz (non‐verbal Montreal Imaging Stress Task) to induce mental stress. Future studies could explore different stress‐induction techniques for a more comprehensive understanding.

• Longitudinal Studies: Longitudinal studies tracking GSS and HRV patterns over time in response to chronic stressors could provide valuable insights into the long‐term impact of stress on autonomic function and assess the effectiveness of the GSS in monitoring stress variations over time. Van Kraaij and colleagues (van Kraaij et al. 2020) found a significant relationship between chronic stress and heart rate over time using various wearable devices. Limitations of the current study are the use of a single‐event intervention to induce mental stress, which may not have activated stress sufficiently in all participants. Future research should consider employing a longitudinal experimental design to examine the impact of accumulating stressors over time. Such an approach would allow for a more comprehensive understanding of mental stress responses and their physiological manifestations. Additionally, our study may have been limited by relatively short Garmin wear periods, as the subjective nature of GSS may require longer monitoring periods to enhance predictive accuracy.

• Additional physiological measures: Our study did not measure or record respiration rate, despite its strong correlation with HRV (García‐González et al. 2000) and its inclusion in the Garmin calculation algorithm (J. Kettunen Saynatsalo (FI); and S. Saalasti Jyvaskyla (FI) 2004). We chose the non‐verbal MIST to minimise artificial fluctuations in respiration rate. However, future studies could benefit from directly monitoring this parameter.

• Additional stress measures: Future studies could incorporate additional physiological stress measures (e.g., cortisol levels) to be compared with the wearable metric and provide a wider picture of the physiological response to the stressor. Also, blood pressure would increase our ability to fully disentangle the sources contributing to stress level estimations and should be considered in future studies. Finally, using an ECG monitor for HRV recording may provide advantages over the IBI‐based monitoring employed in this study, as it allows for the detection and correction of more complex artefacts.

• Moderating variables: Future research should further investigate moderating variables in stress reactivity to deepen understanding of how individual differences shape physiological stress responses, as these factors appear to play a significant role in moderating both the experience and physiological impact of mental stress on individuals. These could include factors such as age, sex, personality traits (e.g., neuroticism or resilience), baseline fitness levels, chronic stress exposure, sleep quality, and lifestyle behaviours (e.g., caffeine consumption or physical activity patterns).

• Chronic Disease Monitoring: Given the association between the stress level score and HRV metrics in the current study, alongside the established evidence of low vagally‐mediated HRV as a marker of health risks (Gidron et al. 2018) (De Couck et al. 2012), it is worth exploring the potential use of Garmin device for monitoring patients with existing health conditions. Continuous, real‐time tracking of psychobiological changes using the Garmin stress score could provide insights into the dynamics of disease progression, including stress responses and comorbidities. Such information may help to better understand underlying physiological processes and serve as an early‐warning system to prevent exacerbation of chronic diseases through timely, tailored interventions.

• Generalisation to a wider population: Our sample primarily consisted of adults—ages 18–39. Although the physiological data suggested a diverse range of participant profiles, future studies could aim to recruit and compare a broader and more representative sample from the general population. It should be noted that a broader age range population could have too much variability as HRV is significantly lower in the older population (Nunan et al. 2010).

Notably, with technological advancements, as well as wider commercial platforms enabling simple assessment of the raw data, these limitations can be reduced. Access to Garmin IBI data and algorithms is becoming increasingly available, facilitating direct examination and comparison of HRV parameters across different wearables, combined with better understanding of the underlying calculations.

5. Conclusion

Our findings underscore the dynamic nature of autonomic nervous system regulation in response to mental stressors, revealing significant variations in physiological responses as function of various external and individual factors. The significant elevation in GSS during the stressful task, coupled with its high association with HR and moderate correlations with certain HRV metrics, suggests that the device effectively detects the physiological response to the mental stress task used in this study.

Elevated HR and sympathetic activity during physiological stress responses, alongside reduced parasympathetic activity indicated by decreased HF power, were observed. GSS mirrored these patterns. Specifically, GSS exhibited a significant increase during stress compared to baseline conditions. Additionally, tonic HRV, exercise, BMI, and sleep duration significantly influenced GSS, with high tonic HRV, regular exercise, and longer sleep duration associated with lower stress scores, while higher BMI and low tonic HRV were linked to higher stress scores. The observed correlations between the GSS and HR and HRV metrics, alongside consistent findings between the pilot and main pre‐registered experiment, support that Garmin Vivosmart 4 mirroring physiological patterns, having the potential for monitoring mental stress level.

Moreover, we found, out of our measured variables, that HR is the most reliable physiological parameter for predicting real‐time self‐reported stress. This finding, in which an easy‐to‐access physiological variable holds substantial potential for monitoring internal states in participants, provides new insights for the field. Importantly, the observation that mental stress, regardless of the stressor, can be detected and predicted through HR may prove valuable across a wide range of applications, mostly for a real‐time assessment or as a biofeedback parameter. However, further studies are needed to examine this finding in more real‐life experimental settings, as well as against other physiological measurements that were not used in our design.

The observed associations highlight the potential of wearable technology, in objectively assessing stress levels in real‐world settings. Their noninvasive and remote nature facilitates longitudinal and ecological research in health, behaviour, and other fields of human research. Accessibility and ease of use enable large‐scale studies, facilitating exploration across diverse populations and contexts. Furthermore, integrating GSS with other physiological and behavioural measurements could provide a comprehensive understanding of stress‐related processes and their impact on behaviour, health, and well‐being.

Future research encompassing a broader range of mental stress induction methods, and longitudinal designs with consideration of individual differences, could further enhance our understanding of the GSS's utility in real‐world stress management, mental and physical alike.

Ethics Statement

Ethical approval was obtained from the ethics committee of Tel Aviv University.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting Information S1

Rosenbach, Hadar , Itzkovitch Alon, Gidron Yori, and Schonberg Tom. 2025. “Assessing Stress Level Scores Against Wearables‐Driven Physiological Measurements,” Stress and Health: e70125. 10.1002/smi.70125.

Data Availability Statement

Experimental data, analysis codes, and task codes are available through the Open Science Framework (OSF). OSF project: https://osf.io/gdr2n.