Validation of heart rate measured by the consumer-level wearable device Fitbit: a cross-sectional observational study

Abstract

[Purpose] The present study aimed to validate the heart rate of the consumer-level wearable device Fitbit Inspire 3. [Participants and Methods] Participants were 22 healthy university students (20.9 ± 0.6 years, 11 males and 11 females). They performed 5-minute resting lying, sitting, walking (3.2 km/h, 4.2 km/h, 6.0 km/h and 3.2 km/h with 3 kg backpack) and running (8.4 km/h running) on a treadmill, and stair climbing, with Fitbit- and electrocardiogram-based heart rate (Fitbit-HR and ECG-HR). The relationship between Fitbit-HR and ECG-HR was evaluated using a generalized linear mixed model. Additionally, Brandt–Altman analysis and absolute percentage error were used to examine the measurement error. [Results] Fitbit-HR was strongly associated with ECG-HR and oxygen uptake. In the Brandt–Altman analysis with Fitbit-HR and reference ECG-HR, most plots were within the limit of agreement. Fitbit-HR showed significant fixed errors, which were, on average, three beats smaller than those of ECG-HR. No significant proportional errors are observed. The mean absolute percentage error for most of the tasks was less than 10%. [Conclusion] The Fitbit Inspire 3 showed a strong correlation and clinically acceptable agreement with ECG-HR in healthy young adults, despite a small, consistent underestimation. Its validity was supported by its association with oxygen uptake.

INTRODUCTION

Chronic pain imposes significant economic burden, making its management and prevention crucial¹⁾. Physical activity and exercise are established methods of managing and preventing chronic pain^{2, 3)}. These activities are often performed under direct professional supervision^4,5,6), which, although valuable, requires individuals to invest resources in accessing on-site facilities. Therefore, telerehabilitation programs are being developed to make chronic pain management more accessible, with initial studies suggesting effectiveness comparable to in-person sessions^{7, 8)}.

Furthermore, recent advancements are aimed at the remote monitoring of heart rate during telerehabilitation by combining consumer-grade wrist-worn devices with information communication technology, thus enhancing safety and effectiveness. The Fitbit^® devices, along with Apple Watch^® and Garmin^®, are among the most commonly used devices⁹⁾. Although the Fitbit has shown good measurement accuracy^{10, 11)}, it tends to underestimate heart rate^{9, 12)}. This tendency to underestimate heart rate, particularly during physical activity, can be attributed to several factors. These include motion artifacts that introduce noise into the sensor data, unstable device positioning and variations in the pressure applied to the skin by the sensor, all of which can lead to measurement inaccuracies¹³⁾. Previous studies have often relied on a single average heart rate measurement for exercise tasks, failing to consider short-term variations or time courses. In addition, the relationship between heart rate and oxygen uptake (VO₂) remains unclear. Addressing these gaps could increase the feasibility of telerehabilitation for remote heart rate monitoring.

We hypothesized that a relationship exists between the Fitbit-based heart rate (Fitbit-HR) and both the electrocardiogram-based heart rate (ECG-HR) and VO₂, with an absolute percentage error (APE) of ˂10%. Therefore, this study aimed to validate the heart rate measurements obtained via Fitbit (Fitbit-HR) in healthy adults.

PARTICIPANTS AND METHODS

This experimental study included a series of resting and exercise tasks: lying and sitting at rest, walking at 3.2 km/h, 4.2 km/h, and 6.0 km/h, walking at 3.2 km/h with a 3 kg backpack, running at 8.4 km/h on a treadmill, and free-paced stair climbing. The stair climbing task involved ascending one floor of 22 stairs, with each step approximately 18 cm in height. These tasks were adapted from Yano et al¹⁴⁾. Each task lasted for 5 min, with no intervals between tasks, during which heart rate and oxygen uptake were continuously measured and synchronized across devices. The study was conducted at Takasaki University of Health and Welfare between October 2023 and June 2024.

The study was performed in accordance with the Declaration of Helsinki and the Strengthening the Reporting of Observational Studies in Epidemiology guidelines. Ethical approval was obtained from the Ethical Review Committee of Takasaki University of Health and Welfare (approval numbers 2241 and 2261). Informed consent was obtained from all the participants.

The participants were healthy university students of the Department of Physical Therapy at Takasaki University of Health and Welfare who were recruited via email. The exclusion criteria were physician-imposed exercise restrictions and illness on the day of the experiment. The sample size for the generalized linear mixed model was not predetermined, as no established guidelines exist. Instead, model adequacy was assessed post hoc using Conditional R² (the combined explanatory power of fixed and random effects).

Reference heart rates were measured using a wireless ECG device (WHS-1; Union Tool Co., Tokyo, Japan) with a sampling frequency of 1,000 Hz. Instantaneous heart rate was calculated at each heartbeat from the RR intervals on a personal computer application, and the post-measurement data were exported.

The Fitbit Inspire 3 (Google LLC, Mountain View, CA, USA) was synchronized with a smartphone and worn on the non-dominant arm of each participant. The device was placed approximately one finger-width proximal to the radial styloid process and tightened to prevent its movement during the tasks. The device utilizes PurePulse technology, a proprietary photoplethysmography (PPG) sensor, to measure heart rate from the wrist. The data were exported from the application via a web interface after each session. The measurement frequency varied, with higher intensities yielding higher frequencies¹⁵⁾. The interval between measurements was 15 s at rest and during low-intensity tasks and 1 s during high-intensity tasks.

Oxygen uptake was measured breath by breath using a mobile gas analyzer (AE 100i; Minato Medical Science Co. Ltd., Osaka, Japan) calibrated prior to each session. The participants wore a tightly sealed mask to ensure accurate measurements of VO₂. The data obtained from three different devices (the ECG device, Fitbit Inspire 3, and mobile gas analyzer) were synchronized using a timestamp recorded on each device. The maximum possible time lag among the three devices was 1 s.

A 30-s moving average was calculated by averaging the values within 15 s before and after each measurement. Afterward, the resulting smoothed data were used to calculate average values by dividing each 3-min period into 1-min segments, excluding the first and last minutes of each 5-min task. In total, 24 time points were generated, with eight tasks per participant. Oxygen uptake was normalized to body weight (VO₂/kg).

Linear mixed models were constructed, with the ECG-HR (model 1) and VO₂/kg (model 2) as dependent variables, the Fitbit-HR as a fixed effect, and participant ID as a random intercept. The Bland–Altman plot was applied to assess the agreement between the Fitbit-HR and ECG-HR, with the 95% confidence interval (CI) of the difference between the ECG-HR and Fitbit-HR (Diff[ECG-HR−Fitbit-HR]) used to evaluate the fixed error (CI excluding 0 indicated fixed error). In addition, a linear mixed model was created with Diff(ECG-HR−Fitbit-HR) as the dependent variable, the mean ECG-HR and Fitbit-HR (Mean[ECG-HR and Fitbit-HR]) as a fixed effect, and participant ID as a random intercept (model 3). A significant coefficient for Mean(ECG-HR and Fitbit-HR) indicated proportional error.

To assess the accuracy of Fitbit-HR in predicting ECG-HR, the APE was calculated using the following formula:

$$\left|\mathrm{ECG}-\mathrm{HR}-\mathrm{Fitbit}-\mathrm{HR}\right|/\mathrm{ECG}-\mathrm{HR}\times 100$$

with mean and median APE values categorized as follows: <10% (high accuracy), 10–20% (good accuracy), 20–50% (reasonable accuracy), and >50% (low accuracy)¹⁶⁾.

All analyses were conducted using R, with the significance level set at α=0.05, and the parameters for the linear mixed models were estimated using restricted maximum likelihood.

RESULTS

As of October 2023, 192 students were enrolled at the Department of Physical Therapy, Faculty of Health and Care, Takasaki University of Health and Welfare, Japan. Of these, 22 (11.5%) volunteered to participate in the study. Participants had a mean age of 20.9 ± 0.6 years (range: 20–22), an equal male-to-female ratio, and a mean body mass index of 21.1 ± 2.1 kg/m² (range: 17.8–25.5). None of the students who chose to participate in the study met the exclusion criteria, and they all completed the experiment; thus, no data were missing from the analysis.

Table 1 summarizes the linear mixed model statistics for ECG-HR (model 1) and VO₂/kg (model 2), with the Fitbit-HR as a fixed effect. The intraclass correlation coefficients (ICCs) for random effects were 0.18 in model 1 and 0.47 in model 2. In both models, the Fitbit-HR showed a strong association with the ECG-HR and VO₂/kg, as indicated by high t-values. The Conditional R² values for models 1 and 2 were 0.868 and 0.774, respectively.

Table 1. Statistics of linear mixed models (Model 1: ECG-HR and Fitbit-HR, Model 2: VO₂/kg and Fitbit-HR).

Random effects		Variance	Standard deviation
Model 1	ID intercept	22.28	4.72
	Residual	101.34	10.07
	ICC	0.18
Model 2	ID intercept	10.40	3.23
	Residual	11.77	3.43
	ICC	0.47
Fixed effects		Estimate	Standard error	t-value	p-value
Model 1	Intercept	6.57	1.97	3.34	0.001
	Fitbit-HR	0.96	0.02	60.89	<0.001
Model 2	Intercept	−15.96	0.90	−17.71	<0.001
	Fitbit-HR	0.29	0.01	53.97	<0.001

ECG-HR: electrocardiogram-based heart rate; Fitbit-HR: Fitbit-based heart rate; VO₂/kg: oxygen consumption per 1 kg; ICC: intraclass correlation coefficient.

The Bland–Altman plots, shown in Fig. 1, depict an upper limit of agreement (LOA) of 24.77 beats/min (95% CI: 23.16–26.38) and a lower LOA of −19.58 beats/min (95% CI: −21.19 to −17.96), with most values falling within these limits. The mean difference between ECG-HR and Fitbit-HR (Diff[ECG-HR−Fitbit-HR]) was 2.60 beats/min (95% CI: 1.65–3.54), indicating a positive fixed error. The coefficient of model 3 for Mean (ECG-HR and Fitbit-HR) was 0.03 (t=1.84, p=0.067), showing no significant proportional error.

Fig. 1.

The Bland–Altman plot of the ECG-HR and Fitbit-HR.

The solid blue center line and shaded area represent the mean difference between the ECG-HR and Fitbit-HR, along with its 95% confidence interval. The upper and lower dashed red lines and shaded areas indicate the upper and lower limits of agreement, along with their 95% confidence intervals.

ECG-HR: electrocardiogram-based heart rate; Fitbit-HR: Fitbit-based heart rate.

The mean APEs for walking at 3.2 km/h and 3.2 km/h with a 3-kg load were 11.21 ± 9.35% and 11.18 ± 10.19%, respectively; however, the median APE was below 10% in both cases. Furthermore, the mean and median APEs for other tasks remained below 10% (Table 2).

Table 2. Mean and median APEs for each task.

Task	Mean APE (%)a	Median APE (%)b
Bed rest	5.48 ± 5.57 (0.10–23.05)	3.62 ± 3.24
Sitting	7.61 ± 6.82 (0.40–41.67)	5.86 ± 3.15
Walk at 3.2 km/h	11.21 ± 9.35 (0.67–41.63)	9.94 ± 5.81
Walk at 4.2 km/h	8.33 ± 6.98 (0.01–27.97)	5.78 ± 5.31
Walk at 6.0 km/h	7.24 ± 5.65 (0.02–27.30)	5.91 ± 2.68
Walk at 3.2 km/h carrying a 3 kg load	11.18 ± 10.19 (0.03–47.62)	8.43 ± 5.27
Running at 8.4 km/h	6.73 ± 6.28 (0.10–29.08)	5.00 ± 4.20
Stair climbing at a self-selected pace	7.29 ± 4.84 (0.27–20.00)	6.67 ± 3.35

^aValues are expressed as means ± standard deviation (minimum–maximum). ^bValues are expressed as medians ± quartile deviation.

APE: absolute percentage error.

DISCUSSION

In the present study, each participant underwent repeated measurements, introducing the possibility of correlations between the data. To account for this, we incorporated participant ID as a random effect in the model to capture participant-specific influences. The fixed effect estimate for the Fitbit-HR was 0.96, with a strong t-value of 60.89, indicating a close association with the ECG-HR. This suggests that a one-beat increase in the Fitbit-HR corresponds to an approximate 0.96-beat increase in the ECG-HR, thereby confirming a robust association. In addition, previous studies have shown a strong association between the Fitbit-HR and ECG-HR during treadmill walking and running¹⁷⁾ and underwater treadmill tasks¹⁸⁾. The low ICC for the random effect (0.18) indicated minimal within-participant variation in the ECG-HR, suggesting that most of the variations were attributable to fixed effects. This finding agrees with the significant fixed effects estimate, which accounts for most of the variance.

Although the linear mixed model showed a strong association between the Fitbit-HR and ECG-HR, it did not sufficiently confirm an agreement between the two measures. Therefore, a Bland–Altman analysis was conducted to assess the agreement. The analysis revealed a positive fixed error of 2.60, indicating a slight underestimation of heart rate by Fitbit, which is consistent with the results of previous studies^{12, 19)}. Although this bias is minor, it can introduce clinically significant errors when prescribing exercises using heart rate-based formulas such as the Karvonen method. Underestimation of the resting heart rate can lead to an overestimation of the heart rate reserve and, consequently, an incorrectly calculated target zone. This issue is compounded by the underestimation of heart rate during the activity itself, which makes real-time intensity monitoring challenging. Therefore, this bias may be acceptable for general fitness tracking in healthy individuals; however, clinicians should account for this systematic underestimation when using Fitbit devices for therapeutic exercise prescription.

The mean and median APEs were below 10% for most tasks, suggesting a high accuracy of the Fitbit-HR in estimating the ECG-HR; this finding agrees with previous studies that reported a general mean APE under 10%²⁰⁾ and comparable accuracy in older adults²¹⁾. However, for slower tasks (3.2 km/h walking and 3.2 km/h walking with a 3-kg load), the mean APE slightly exceeded 10%, likely due to the Fitbit’s lower sampling rate during resting and slow-walking tasks (intervals of up to 15 s). Such low- frequency sampling not only contributes to the overall error, but also limits the ability of the device to track dynamic heart rate changes, such as rapid heart rate dynamics at the beginning or end of exercise. This issue, in addition to the tendency for pulse wave measurements to be sensitive to motion artifacts²²⁾, may have contributed to higher outlier effects in these tasks. Although the median APE was <10%, the increased error at this low intensity warrants clinical consideration. Low-intensity walking is the basis for many physical therapy interventions, including early cardiac rehabilitation, postoperative weaning, and exercise for frail older adults. In these situations, heart rate is a key indicator for assessing physiological tolerance and ensuring patient safety. More attention is required when monitoring patients during low-intensity activity in telerehabilitation, as this measurement error can mask meaningful heart rate responses to activity.

Further research is warranted to address the variability in measurement accuracy in different conditions. Conflicting evidence exists regarding accuracy at varying exercise intensities. Although some studies report a higher APE at lower intensities²³⁾, others indicate reduced accuracy at moderate intensities^{24, 25)}.

In a linear mixed model with VO₂/kg as the dependent variable, the Fitbit-HR as a fixed effect, and participant ID as a random effect, the ICC for the random effect was 0.47 and the t-value for the Fitbit-HR was 53.97, indicating a strong association. The model’s Conditional R² of 0.774 suggests a high explanatory power.

The relatively high ICC indicates that VO₂/kg values were consistent within participants, with substantial differences observed between participants. Such differences may result from physiological variations, as men with greater muscle mass typically exhibit a higher VO₂/kg than women²⁶⁾. Similarly, improvements in energy efficiency, achieved through changes in body composition from running and dietary programs, affects VO₂/kg²⁷⁾. These factors underscore the fact that individual differences in body composition and efficiency may account for much of the variability in VO₂/kg across participants.

Despite this between-individual variability, the Fitbit-HR as a fixed effect remained strongly associated with VO₂/kg, confirming a close relationship between heart rate and oxygen uptake as measured by the Fitbit. This trend is consistent with the Fick equation, which shows that cardiac output, calculated as stroke volume multiplied by heart rate, plays a key role in determining oxygen uptake²⁸⁾. In addition, the relationship between the ECG-HR and VO is well documented²⁹⁾.

The findings of this study reinforce the validity of the Fitbit-HR as an estimator of VO₂/kg, as it effectively predicted oxygen uptake during walking and running. Moreover, this supports the utility of the Fitbit as a viable tool for monitoring heart rate and estimating associated physiological metrics in controlled activity settings.

This study applied a linear mixed model to repeated-measures data, a method suitable for addressing the inherent lack of independence within repeated measures, as each participant contributed multiple data points. By incorporating inter-participant and inter-measurement variabilities as random effects, this model enabled robust association estimates between the Fitbit-HR, the ECG-HR, and VO₂/kg.

However, three main limitations restrict the generalizability of the study findings. First, the findings of this study are based on a homogeneous sample of healthy young university students. The generalizability of these findings to typical physical therapy target populations, such as older adults or individuals with chronic diseases, is limited. Decreased skin elasticity and wrinkles associated with aging³⁰⁾ can compromise the fit of a Fitbit device; reduced capillary blood flow³¹⁾ may hinder heart rate measurements based on the principle of photoplethysmography. Although the accuracy of heart rate measurement is relatively high in older adults²¹⁾, clinicians should be aware of these factors. For patients with cardiovascular conditions such as atrial fibrillation, the resulting arrhythmia can compromise the accuracy of heart rate measurements³²⁾. In addition, darker skin tones and higher body mass index increase measurement error³³⁾. Therefore, it should be recognized that clinical populations may exhibit a higher measurement error than we have observed. Validation studies are warranted to determine the accuracy of this device in these specific clinical populations. Second, although this study validated the Fitbit-HR under controlled treadmill conditions, its validity in free-living settings remains uncertain. Challenges in detecting moderate- to high-intensity activity based solely on heart rate in such settings¹⁵⁾ warrant further research to assess the accuracy of Fitbit across different conditions, with potential applications in telerehabilitation. Third, the data obtained from the three different devices (the ECG device, Fitbit Inspire 3, and mobile gas analyzer) were synchronized using a timestamp recorded on each device. This synchronization resulted in a potential time lag of up to 1 s between the devices. Although the effect of a 1-s time lag was negligible in our study, as the mean data from 180-s intervals during each task were used for comparisons, our validation results may not apply to data recordings that require high time resolution.

Conference presentation

The main part of this research was presented at the 29th Japanese Society of Physical Therapy Fundamentals Congress.

Authorship contribution

Daisuke Higuchi: Conceptualization, Methodology, Writing − Original Draft. Yuko Takahashi: Methodology. Shigeya Tanaka: Methodology. Yosuke Tomita: Supervision, Writing − Review & Editing

Funding

This work was supported by JSPS KAKENHI (Grant Number JP24K14179).

Conflict of interest

The authors have no conflicts of interest to declare that are relevant to the content of this article.