Cardiorespiratory fitness (CRF) is an important component of physical fitness, representing the body’s ability to take oxygen in and deliver this oxygen to muscle cells throughout the body during physical activity. Previous studies have emphasized the importance of CRF, providing convincing evidence that CRF is closely related to all-cause mortality [1,2]. In addition, CRF is known to be correlated with various physiological factors, such as body composition and blood pressure, and psychological factors, such as depression [1-4]. Erikssen et al [5] have reported that a change in physical fitness is a strong predictor of mortality. They found that a small improvement in physical fitness can significantly lower the risk of death.

Maximal oxygen uptake (VO₂ max) is regarded as a representative feature of CRF. The current gold standard for measuring VO₂ max is a metabolic gas analysis during a maximal graded exercise test (GXT) on a treadmill or other equipment, such as cycle ergometer. Even though the maximal exercise test provides an accurate measurement of VO₂ max, there are several limitations. The maximal exercise requires a high level of motivation from the subject and should be performed under medical supervision for older or high-risk subjects who need this test the most [6]. Furthermore, the gas analysis requires expensive equipment and a trained technician to operate the process [7]. In addition, because of the high cost and inconvenience, it is impractical to repeat the maximal exercise test to regularly monitor VO₂ max.

Several estimation models have been developed to estimate VO₂ max. Some of these models have developed a submaximal exercise protocol in an attempt to overcome the limitation of the maximal test [8-10]. Submaximal models obtain exercise-related data through a specified exercise protocol, such as shuttle run, and build estimation models along with other anthropometric features. Although submaximal models have overcome some of the limitations of the GXT, they still require trained personnel to conduct the submaximal test, and familiarity with the exercise protocol could affect the results of the test [11], making it unsuitable for regular VO₂ max monitoring. There are other estimation models that do not involve an exercise protocol [12,13]. These models estimate VO₂ max by collecting data from physical activity and heart rate (HR) from daily life and using the relationship between the collected data and VO₂ max. Although these methods are more suitable for regular VO₂ max measurement, they are time-consuming (requiring a week of data collection) and use multiple devices, making it uncomfortable for application in daily life.

In our previous study [14], we developed a nonexercise VO₂ max estimation model using a new feature, maximum activity energy expenditure (aEEmax), which was calculated using activity energy expenditure and HR. Using aEEmax, we were able to build an accurate estimation model. However, aEEmax and our previous model were validated only in homogenous subjects, young Asian males. Furthermore, the device was worn on the chest, which might cause discomfort when used in daily life.

The aim of this study was to overcome the limitations of our previous study by using a wristwatch-type fitness tracker with various groups in terms of age and sex. We also sought to develop a new VO₂ max estimation model using aEEmax and the slope between physical activity and HR as new features, which could be applied to daily life data collected from a single convenient device worn on a wrist with a relatively short estimation time.

The general characteristics of all subjects are summarized in Table 1. The average value for age, weight, height, and percent body fat for excluded male subjects were 42.0, 72.8 kg, 171.0 cm, and 25.0% respectively. For excluded female subjects, the average values were 41.7, 56.5 kg, 158.4 cm, and 31.6%, respectively. A student t test was performed to compare the P value for age, weight, height, and percent body fat between included and excluded subjects. For male subjects, the P values were .646, .937, .883, and .026, respectively. For female subjects, the P values were .936, .415, .663, and .497 respectively.

The Pearson correlations between VO₂ max and selected features are shown in Table 2. The Pearson correlations between selected features are also summarized. The correlations between VO₂ max and independent variables were all statistically significant (P<.001 for all). The independent variable that showed the highest correlation was sex, with a correlation of .675. The lowest correlation for the model was the slope, with a correlation of .237.

The multiple linear regression analysis for the model is shown in Table 3. The scatter plot for measured VO₂ max versus predicted VO₂ max is shown in Figure 1. The R² for the model was 0.651, and the SEE was 3.518 mL/kg/min. As shown in Table 3, the decrease in R² and the increase in SEE were small for the cross-validation result of the PRESS method. R² decreased by 0.032 and SEE increased by 0.148 mL/kg/min. The scatter plot of the multiple linear regression for the model is shown in Figure 2.

The model was further validated by dividing the groups into various subgroups and calculating the CE and SD. Each subgroup was divided into 2 groups according to age, sex, and measured VO₂ max. The results are shown in Table 4. The CEs were positive for the younger group and negative for the older groups. The CE for the younger group was 0.024 mL/kg/min. The CE for the older group was −0.021 mL/kg/min. The CEs were all positive for the subgroups divided according to sex. The CE was positive for individuals with high VO₂ max and negative for individuals with low VO₂ max. However, as shown in Table 4, CE values were low for all subgroups, indicating that our model does not overestimate or underestimate VO₂ max.

In this study, we developed VO₂ max estimation models with data collected from the commercially available device, Fitbit, which could estimate VO₂ max conveniently from daily life. The commercial device used in this study provided data on physical activity, along with its physiological response (HR). Even though some previous studies have reported that Fitbit does not always generate accurate data [19,20], our method for estimating VO₂ max does not depend on the absolute value of each data point. It rather depends on the trend of a large set of data points and would balance the inaccuracy of a single data point. Even though Fitbit data were biased under certain conditions, a large number of data collected from daily life would minimize this bias. This characteristic of our 2 novel features would only need calibration to be applied with other hardware and processing methods. Previous studies reported that lower HR was not linearly correlated with physical activity [16]. It is also known that heterogeneous recovery of HR after physical exercise does not have a clear correlation with VO₂ max [21]. Therefore, in an attempt to select the data during physical activity, we have selected periods of data where HR is greater than 120 beats/min and where both HR and physical activity increased at the same time to calculate aEEmax and the slope. By calculating 2 features from the linear relationship between HR and energy expenditure, our model would estimate VO2max accurately without any protocol or training.

In our previous study [14], aEEmax and our estimation model were validated with a homogenous group of young healthy Asian males using aEEmax and BMI (VO2max=0.192 x aEEmax – 0.708 x BMI). In this study, we have validated aEEmax and the model with a large pool of subjects, including both men and women aged from 20 to 65 years. The validation of the model was performed by cross-validation with the PRESS method and by calculating the CE of the subgroups. The cross-validated result shows that the validation sample fitted well with the model with little error. Furthermore, the CE for the model shows that our methodology did not significantly overestimate nor underestimate VO₂ max for all subgroups, whereas other studies [12,22] have reported significant overestimation and underestimation with both highly and poorly fitted individuals.

To the best of our knowledge, this study is the first to provide a VO₂ max estimation model with a commercially available device on a wrist without any specific protocol. Tönis et al used a submaximal exercise protocol to estimate VO₂ max [23]. Polar Electro Oy Inc developed a nonexercise protocol, Polar Fitness Test, and devices to estimate VO₂ max. However, Esco et al [24] reported that the Polar Fitness Test had low accuracy for estimating VO₂ max when it was tested with one of its own products. Altini et al [25] developed a nonexercise estimation model for VO₂ max with data collected from daily life; however, they used an ECG necklace with a wet electrode attached to the chest and stomach, which could cause an inconvenience when used in daily life. Our method allows individuals to measure their VO₂ max on a wrist without the requirement for any electrode attachment. In addition, although other smartwatches, including Fitbit, require a specific protocol, such as running for at least 10 min on flat terrain, our protocol does not require any protocol and allows for easy monitoring of physical fitness.

Instead of using the absolute value of HR or physical activity data, we developed 2 new features to portray physical fitness. The change in HR for a given physical activity differs depending on the physical fitness of a subject [8]. Thus, the slope between HR and physical activity would be smaller and aEEmax would be larger for a subject with higher VO₂ max. As shown in Table 2, aEEmax has a moderate correlation with VO₂ max, supporting our hypothesis. Features based on physical activity can easily fluctuate depending on change in the short-term lifestyle of the subject during the period of data collection. Those features would be vulnerable to a sudden increase or decrease in the amount of physical activity. On the contrary, aEEmax and the slope represent the relationship between physical activity and HR and thus would be less affected by a sudden short-term change in physical activity. In addition, aEEmax and slope have been shown to be applicable to a wide range of subject ages and different genders. Other studies [26-28] have been validated with a relatively homogenous group compared with this study. The percent body fat used in this study was obtained using a professional bioimpedance analyzer. However, we have used general percent body fat which did not necessarily require a professional analyzer. There are products available, such as an AURA device, and ongoing studies about measuring body fat percentage from the wrist [29]. These efforts will make the measurement of percent body fat more accessible to the public.

There are limitations to this study. First, our methodology needs to be validated with more devices worn on a wrist. There are many commercially available devices that provide physical activity and HR data. To provide a more generalized VO₂ max estimation method, it is important to prove device independency of our method. In addition, this study was conducted with healthy subjects who were not taking any medication that might affect the HR. Another limitation of this study was error with maximal HR calculated from a basic population-derived formula. A more accurate method for calculating maximal HR could increase the accuracy of our model. A future study could include subjects who are on cardiac-related medication. Additionally, in our previous study, we have shown that 900 min of data were enough to calculate aEEmax; however, it would be worthwhile to observe change in the correlation coefficient of aEEmax for a longer period of time.

In summary, we have developed a new estimation model for VO₂ max using novel features, aEEmax and the slope between physical activity and HR, along with other anthropometric variables. The new features represent the relationship between physical activity and its physiological response. The high correlation between VO₂ max and aEEmax is in agreement with our previous study and supports our hypothesis. Our model requires data from only 3 days of daily life, without any specific exercise protocol. This hypothesis was validated with a diverse and large number of participants based on age and sex. Furthermore, all material required for our study is available in the conventional market as fully built products. The result of this study allows individuals to measure their VO₂ max conveniently in their daily life without any burden of an exercise protocol and allows them to easily monitor physical fitness.