This is a cross-sectional study, within‐subject design to compare the HR accuracy from the GW6 with criterion measures of HR (Polar H10 chest strap) during a maximal ramp CPET conducted on a treadmill. All participants had their health conditions evaluated by a physician before the tests. Participants were instructed to abstain from alcoholic beverages and any medication for 24 hours, to avoid caffeine intake for 12 hours before the test, and to hydrate with water only. Body weight and height were measured using a weighing scale and stadiometer (InBody 770 analyzer), and skin type was measured via the Fitzpatrick skin type scale [6]. The environment in the laboratory was standardized throughout all sessions by maintaining an ambient temperature of approximately 22°C. All technical procedures followed the American College of Sports Medicine guidelines [7]. We followed the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement: guidelines for reporting observational studies [8].

The inclusion criteria were healthy, recreationally active participants who engaged in at least 150 minutes of moderate-intensity activities and/or 75 minutes of vigorous activities per week.

Exclusion criteria were overweight and obesity, recent musculoskeletal injuries, any cardiovascular disease, and use of medications or supplements (eg, beta-blockers and caffeine) that impact HR measures. Participants signed the informed consent form after a verbal and written explanation of the activities to be carried out and a full understanding of the possible risks involved in participating in the study.

Initially, 76 participants began the CPET protocol. Four were excluded because of execution failures, and 17 were excluded because of device failures. The participant selection flowchart can be found in the Results section. Execution failure refers to problems that occurred before the trial started. Examples include the device not being set up correctly; failure to pair or synchronize the device; and incorrect entry of participant information into the device. Device failure refers to problems that arose during or after the trial. Examples include: loss of connection during CPET execution; the Polar H10 chest strap producing an unstable, missing, or abruptly dropping HR signal for any portion of the test; and devices that were correctly configured but still failed to generate all required output files.

All participants performed maximal ramp CPET on a motorized treadmill (COSMED T170 DE) to determine HR at increasing intensities and maximal oxygen consumption (determined as the highest value observed during the test). The test was started at 5 km∙h^-1 and the speed increased by 1 km∙h^-1 every minute, with the slope kept constant at 1% [9] until they met at least two out of the following three conditions: (1) reached 90% of their maximal HR value calculated using Tanaka’s equation [10], (2) achieved a respiratory exchange ratio (RER) greater than or equal to 1.05 [11], and (3) reached maximal voluntary exhaustion, evaluated by the Borg (6-20) subjective scale of effort [12]. Gas exchange recording was performed using the K5 portable metabolic analyzer (COSMED). Device setup, calibration, and data extraction were conducted via COSMED OMNIA Metabolic software. The equipment was calibrated before each test as recommended by the manufacturer.

The test was terminated according to the termination criteria recommended by the American College of Sports Medicine guidelines [13], which include the occurrence of cardiovascular, respiratory, or musculoskeletal symptoms, abnormal blood pressure responses, significant rhythm disturbances, or voluntary exhaustion. All tests were supervised by a physician trained in CPET.

During the test, participants wore a chest strap HR monitor (Polar H10) and the GW6 smartwatch. HR data were evaluated across 5 intensity zones commonly used in various applications (50%‐60%, 60%‐70%, 70%‐80%, 80%‐90%, and 90%‐100% of HRmax). The average HR was calculated for each of the 5 intensity zones based on the maximal HR measured by the Polar H10 from the whole test. This approach is consistent with standard clinical practice in cardiology and incremental exercise testing and was chosen to provide a stable and representative estimate of the physiological response at each workload while minimizing the influence of beat-to-beat variability and transient signal artifacts.

Resting HR was measured using the Polar device before the start of the V̇O₂ max test protocol. Participants were seated quietly for several minutes to ensure a stable baseline prior to data collection.

The Polar H10 chest strap (Polar Electro Oy) served as the criterion method for measuring HR. This device has previously been validated against ECG [14]. Each participant was fitted with the Polar chest strap, set to collect continuous HR data throughout the protocol. On completion of the activity, data were transferred from each Polar transmitter to a central laptop to be viewed and analyzed using Polar Flow (version 6.25.2, Polar Electro Oy).

The GW6 (Samsung Electronics Co Ltd) is a multisensor smartwatch. Accelerometer, gyroscope, barometer, magnetometer, and photoplethysmography are some of the GW6’s sensors. During our experiments, all GW6s were connected to separate smartphones (Galaxy A22, Samsung Electronics Co Ltd), and all HR data were exported using the Samsung Health app. In cases where the activity data needed to be trimmed, full health data were exported via the Samsung Health app, with the raw data CSV file processed using Python (version 3.12.7; Python Software Foundation).

Participant height, date of birth, and wrist orientation were entered into each device before the protocol. The GW6 was positioned on the left arm located just above the wrist, as per manufacturer instructions (2 cm above styloid process of radius). In addition, the GW6 was set to record the activity protocol as a workout, being started and stopped at the commencement and completion of the activity. The GW6 was set to record “treadmill run,” as this was the most relevant option available on the watches.

Sample size calculation was performed on a web-based calculator according to Arifin [15]. The lowest acceptable ICC was 0.50, and the expected reliability was ICC=0.75 [16]. The number of raters or repetitions considered was 2 (Polar H10 vs GW6). A significance level of α=.05 (2-tailed) and a power of 80% were set. As a result, 45 participants were needed for this study. However, considering a dropout rate of ~15%, we recruited 55 healthy adults, composed of 30 men and 25 women, aged between 18 and 59 years.

All data were analyzed using a custom-made Microsoft Excel spreadsheet [17]. Data are shown as mean (SD) or median (IQR). The validity of the wearable was determined by the following statistical tests: (1) Correlation between the wearable and the criterion measurement: ICC, common cut-off points for validity assessment: poor (<0.50), moderate (0.50 to 0.75), good (0.75 to 0.90) and excellent (>0.90) [16]; (2) Pearson correlation coefficient; (3) Systematic differences between the wearable and the criterion measurement: median absolute error (MAE) and median absolute percentage error (MAPE) compared to the criterion measurement (criterion-wearable)/criterion × 100). The MAPE was considered with acceptable error when <10% [18]. Additionally, root mean square error (RMSE) was calculated. A predefined subgroup analysis was conducted by stratifying participants according to skin tone using the Fitzpatrick classification. To preserve statistical power, Fitzpatrick categories were aggregated into 3 groups based on pigmentation level. Device performance was then evaluated within each group using the same accuracy and agreement metrics across exercise intensity zones. This subanalysis aimed to assess whether skin tone influenced HR estimation during the incremental cardiopulmonary exercise test.

Bland-Altman analysis was performed to quantify the mean bias and 95% limits of agreement (LoA) between the GW6 and the Polar H10 across the full dataset and within each predefined HR zone (50%‐60%, 60%‐70%, 70%‐80%, 80%‐90%, and 90%‐100% HRmax). For each comparison, the mean difference (device − reference), the corresponding SD, and the upper and lower LoA (mean difference, 1.96 SD) were calculated. Visual inspection was used to assess whether the differences were homoscedastic and whether any proportional bias was present across the range of HR values.

To further characterize agreement between the GW6 and the Polar H10 reference, we computed the coefficient of variation (CV%) for each intensity zone as CV%=(SD/mean) × 100, applying the formula separately to the GW6 and H10 HR values. Additionally, 2-way repeated measures ANOVA was performed to examine whether HR values obtained from GW6 differed from those recorded with the Polar H10 across the 5 predefined intensity zones (% HRmax).

The study protocol was approved by the Investiga Instituto de Pesquisa Ethics Committee (CAAE 62031722.4.0000.5599), and all procedures were conducted following the standards set by the Declaration of Helsinki. Data were anonymized before analysis to ensure confidentiality. No monetary or other forms of compensation were provided. The manuscript contains no identifiable information or images of individual participants. All participants were informed about the study procedures and provided written informed consent.

The primary aim of this study was to evaluate whether the GW6 provides HR estimates that are in at least moderate agreement with a criterion reference chest-strap (Polar H10) across 5 intensity zones (Z1-Z5) of a maximal ramp CPET. The data partially support our hypothesis. Moderate agreement was observed at lower intensities (50%‐60% HRmax), improving to good agreement at moderate intensities (60%‐70% HRmax). At vigorous intensities (70%‐80% and 80%‐90% HRmax), ICC values declined, before increasing again at maximal effort (90%‐100% HRmax), where agreement reached good-to-excellent levels. These findings indicate that between-subject agreement is not monotonic across exercise intensity and is influenced by changes in HR variability across zones. Importantly, despite fluctuations in ICC, absolute and relative error metrics demonstrated a consistent improvement with increasing intensity. MAE remained low across all zones (approximately 1‐3 bpm), while MAPE decreased progressively from 2.9% at low intensity to approximately 0.6% at vigorous-to-maximal intensities. These values are well within thresholds generally considered clinically acceptable for HR monitoring during exercise.

A “U-shaped” pattern in accuracy was observed, characterized by a decline in reliability (as measured by ICC) at intermediate workloads (70%‐80%), followed by a partial recovery at near-maximal intensities. This pattern reflects changes in the consistency of participant rankings (ICC) rather than absolute accuracy. While the ICC suggests reduced consistency in participant rankings at moderate to vigorous intensities, the absolute error metrics (MAE and MAPE) demonstrate a linear improvement in accuracy as intensity increased. The decline in reliability at intermediate workloads is likely attributed to transient mechanical artifacts, such as increased variability and amplitude of arm motion, which may interfere with optical signal stability. Conversely, near-maximal effort, arm movement often becomes more rhythmically constrained, which may help stabilize the photoplethysmography signal and partially restore reliability. This pattern highlights the influence of workload-specific mechanical factors on the observed changes in reliability, rather than contradicting the general trend of increased physiological noise at higher intensities.

To date, to the best of our knowledge, only our study and the study of Abt et al [4] used CPET during HR measurement in healthy volunteers. When comparing our findings with those reported by Abt et al, both studies demonstrate good agreement between smartwatch-derived and criterion HR values, although the magnitude of validity differs according to exercise intensity and device model. Abt et al, with 15 participants, found very high correlations (r=0.87‐0.98) and trivial bias for HRmax measurements obtained with the Apple Watch compared with a Polar reference, indicating strong validity at maximal effort. In contrast, our study, which included 55 participants, showed moderate to good reliability for the GW6 across exercise intensities, with the lowest agreement observed in the 80%‐90% HRmax range. Nevertheless, the MAPE remained small, supporting acceptable accuracy for practical use during exercise. Overall, the larger sample size of our study reinforces the consistency of these observations and suggests that, while optical sensors in current smartwatch models provide increasingly valid HR estimates, their performance may still vary across intensity domains and device types, particularly at submaximal workloads where motion artifacts are more likely.

In this study, the relative reliability (ICC) values obtained in the 5 HR zones are lower compared to those recently measured in individuals with coronary artery disease [5]. Kim et al [5] used an earlier-generation Samsung device (Galaxy Watch 4) and reported values above 0.90. The ICC values in this study ranged from moderate to excellent depending on exercise intensity, with the highest agreement observed at higher intensities (eg, 90‐100% HR zone), while lower-to-moderate ICCs were observed in the remaining zones. These discrepancies can be partially explained by differences in study designs. Kim et al [5] evaluated individuals with coronary artery disease and used the modified Bruce protocol, while in this study, we selected healthy individuals and used a running exercise test protocol with 1-minute stages with constant inclination. In addition, the higher ICC observed by Kim et al [5] may be due to participants being instructed to hold a handrail in front of them as a safety measure during the CPET. As the hand position was fixed during the test, generating less movement artifacts, HR accuracy may have been overestimated. Additionally, with the aim of determining the accuracy of wearable devices (among them, the Samsung Gear S) to measure HR at rest (in the supine and seated positions) and during exercise (walking and running on a treadmill and cycling on a cycle ergometer) in healthy individuals, Wallen et al [19] reported ICC values for HR of 0.80 (0.40 to 0.93). Our results are similar if we consider the 95% CIs.

The MAE measurement of the Samsung Galaxy Watch 4, as reported by Lima et al [20] under hands-free treadmill conditions (2.92 bpm), aligns closely with our findings, which demonstrated a range of 1.00 to 3.00 bpm, depending on the intensity zone. In addition, a coefficient of variation ≤10% was arbitrarily suggested as the standard criterion used to establish an acceptable level of reliability [21]. We consider acceptable accuracy values to be between 5%‐10%.

Aiming to investigate the HR measurement accuracy of 3 wrist-worn devices (Apple Watch 6, Polar Vantage V, and Fitbit Sense) during various activities (sitting, walking, running, resistance exercises, and cycling), Hajj-Boutros et al [22] showed that Apple Watch 6 appears to have the highest level of accuracy for HR monitoring for all 5 activities, followed by Polar Vantage V and Fitbit Sense. However, these investigations were conducted using task-based protocols with fixed intensities, which differ substantially from the physiological and biomechanical demands of a continuous incremental CPET. This study extends this body of evidence by evaluating HR accuracy across predefined intensity zones during a maximal ramp CPET, a context characterized by progressively increasing cardiovascular load, changing movement dynamics, and greater susceptibility to motion-related artifacts. This approach provides complementary and clinically relevant information that is not captured by activity-based protocols, particularly regarding device performance under conditions of rapid HR transitions and high exercise intensities.

In contrast to our findings, Thomson et al [3] reported that MAPE increased with exercise intensity during a Bruce protocol test, which consisted of 3-minute stages with progressively increasing speed and incline until voluntary fatigue. Heart rate reserve (HRR) was categorized into 5 intensity zones: very light (<20% HRR), light (20%‐40% HRR), moderate (40%‐60% HRR), vigorous (60%‐85% HRR), and very vigorous (>85% HRR). Their results showed MAPE increasing from ~2.4% at the lowest intensity to ~4.7% at the highest [3]. In this study, we found a reduction in MAPE with increasing intensity. These discrepancies can be partially explained by the differences in maximal tests (Bruce vs 1-min stages and constant incline), smartwatch models (Apple Watch vs Galaxy Watch), and handrail use during testing (hands-on vs hands-free), and possibly differences in skin color type, which could influence photoplethysmography readings. Although our subgroup analyses stratified by Fitzpatrick skin type did not reveal a consistent pattern of worsening performance across intensity zones, some differences in ICC estimates were observed between skin tone groups in specific zones. However, these differences should be interpreted with caution, given the limited sample size within subgroups.

Our findings suggest that the GW6 can provide reasonably reliable HR data during a maximal treadmill test, which may be of interest to clinicians, exercise-science practitioners, and active individuals who seek a convenient alternative to laboratory-grade equipment. The device should be regarded as a supplemental tool rather than a replacement for ECG-based measurements when precise rhythm analysis or diagnostic accuracy is required. In settings where cost or accessibility limits the use of more expensive monitors, GW6 offers a simple, affordable option for estimating HR, provided its limitations at near-maximal workloads are acknowledged. The use of stage-averaged HR in our study aligns with common practice in clinical exercise testing and CPET, facilitating physiological interpretation and comparability with previous studies. Although alternative approaches, such as shorter averaging windows or peak values within stages, may better capture rapid signal fluctuations, they may also be more susceptible to transient noise, particularly in wearable-derived measurements. Future studies specifically designed to compare different HR aggregation methods may help determine whether such alternatives provide additional value for device validation across different exercise intensities.

An operational aspect that emerged during data collection was the device failure rate of 22% (Figure 1), which reflects instances where the device did not generate usable HR data despite correct placement and adherence to protocol procedures. Although the focus of the study was accuracy during valid measurements, this level of attrition has practical implications. A failure rate of this magnitude may hinder consistent data acquisition in real-world laboratory or training environments, reducing the feasibility of using the device as a reliable continuous monitoring tool. This observation reinforces the need for device-specific robustness improvements and highlights the importance of considering both accuracy and operational reliability when evaluating wearable sensors.

Some limitations of this study should be highlighted. The devices were assessed in young, healthy participants exercising in a standardized laboratory setting. Therefore, our study can only be generalized for this population and setting only. GW6 was placed in a standard position, not considering the effect of left or right side on the accuracy of HR measurement. Therefore, future studies could consider testing watch side randomization to assess HR accuracy. HR measurement was performed using the Polar H10 chest strap, which has a good correlation with ECG. However, this should be considered a limitation.

Every test was stopped only after 2 objective or subjective signs of near-maximal effort were present. For the participants who terminated the test after satisfying the HR and RER criteria without reaching Borg>18, the “90%‐100% intensity zone” may correspond to a submaximal range (eg, ≈81%‐90% of true physiological HRmax). This could slightly overestimate device accuracy because the devices were not challenged at the absolute physiological limits of those individuals.

This study evaluated the validity of GW6 only during a maximal ramp test performed on a treadmill. The continuous, relatively rhythmic arm swing that occurs while running may differ substantially from the movement patterns seen in other common training modalities such as resistance training, high-intensity interval training, or cycling, where wrist motion can be more erratic, static, or intermittent. Therefore, the performance of the embedded photoplethysmography algorithm observed in this investigation may not be directly applicable to those activities. Readers are cautioned against generalizing these findings beyond treadmill-based exercise, and future work should examine accuracy across a broader range of exercise modalities.

The study encountered a 22% device failure rate, leading to the exclusion of 17 participants due to unusable HR data. While the primary aim was to investigate accuracy during valid recordings, this attrition rate represents a limitation of the operational performance of the device. Such failures may impact the practical use in laboratory or field settings, and future research should investigate underlying causes and potential technical improvements to enhance data reliability. It should be noted that if device failure is correlated with specific physiological characteristics (eg, darker skin tone or specific arm movements), the remaining valid sample (n=55) may represent a “best-case” scenario, potentially inflating the accuracy estimates relative to the general population. However, the subgroup analyses stratified by skin tone according to the Fitzpatrick classification did not demonstrate statistically or clinically meaningful differences in device performance.

Additionally, we analyzed a potential survivability bias to identify any patterns among participants excluded. Initially, we conducted an empirical analysis of demographic variables between the failure group and the valid group, focusing on BMI, age, and skin pigmentation. No significant patterns were found in these variables. A second analysis was conducted to address concerns about failed trials from female participants, under the hypothesis that smaller wrist sizes might have affected the proper accommodation of smartwatches. However, this analysis also revealed no discernible patterns. While no bias was identified in the initial analysis, further investigations may be necessary.

This study indicates that the GW6 photoplethysmography-based technology demonstrated reasonable validity in monitoring HR during a maximal cardiopulmonary exercise test in healthy participants.