Introduction

VO₂ max, introduced by British physiologists Hill and Lupton, is the maximum amount of oxygen an individual can utilise during exercise [1]. It provides an evaluation of cardiorespiratory fitness and has been established as an important barometer of health [2–4]. Epidemiological evidence supports an inverse and independent association between cardiorespiratory fitness and mortality, as measured by VO₂ max or metabolic equivalent (MET) [5–7]. It is a predictor of cardiovascular events [8–10], and some studies report it to be more predictive of cardiovascular and all-cause mortality than well-known risk factors such as hypertension, obesity, and hypercholesterolemia [11–13]. In recognition of its significance, the American Heart Association advocates for the routine assessment of cardiorespiratory fitness and has proposed it as a clinical vital sign [14].

VO₂ max is the most widely accepted measure of cardiorespiratory fitness, and the gold standard for its assessment is indirect calorimetry [15]. Indirect calorimetry measures maximal oxygen uptake by recording oxygen consumption and carbon dioxide production during cardiopulmonary exercise testing (CPET) [16]. While it is the most accurate form of measurement, it is costly and resource-intensive, limiting its use in clinical practice and among the general population [17]. To overcome these difficulties, VO₂ max is often derived from submaximal exercise tests that require less time and are considered lower-risk [18]. Error for VO₂ max predictions from submaximal tests is reported to range between 1.6 and 4.1 mL/kg/min [19,20].

Wearable devices also use submaximal exercise to derive VO₂ max [17], and their recent proliferation has democratised monitoring of physiological measurements. They provide unobtrusive, longitudinal monitoring outside the confines of exercise physiology laboratories or healthcare settings, and may facilitate remote monitoring of cardiorespiratory fitness [21,22]. The widespread uptake of wearable devices has been catalysed by the ready availability of a suite of health-related metrics, and their increasing prominence is illustrated by the American College of Sports Medicine (ACSM), who identified wearable technology as the number one fitness trend for 2025 [23]. Over half the population in many countries now owns a wearable device – ranging from smartwatches to smart rings – with global user numbers expected to grow significantly in the coming years [24]. The wearable technology market is projected to reach a valuation of $186 billion by 2030 [25], and Apple Watch holds the largest market share with over 100 million users worldwide [26].

The first-generation Apple Watch was released in 2015, and its VO₂ max estimation feature was introduced with watchOS 4 in September 2017 [27]. This enabled Apple Watch Series 3 and later to estimate VO₂ max using data such as heart rate, Global Navigation Satellite System-derived metrics (e.g., speed, distance, and elevation), and user demographics such as sex, age, height, and weight [28]. Apple subsequently updated its VO₂ max algorithm in 2021 as part of watchOS 7 [28]. While this feature evoked the prospect of an assessment method that could be conducted independently of healthcare or exercise professionals, it was accompanied by notable challenges.

The accuracy of wearable-derived VO₂ max estimates represents one of the most significant challenges. A multitude of factors can influence measurements: physiological factors such as heart rate response to exercise and perfusion; environmental factors like skin contact and motion [29]; or external influences, including caffeine intake and carrying additional weight beyond body weight, such as bags or equipment [28]. Compounding these issues is the proprietary and evolving nature of the algorithms used by wearable devices. With each new hardware or software update, algorithms are subject to change, necessitating recurrent validation by researchers [18,30,31]. This validation is hindered by the protracted process of conducting and publishing research. Consequently, fewer than 5% of consumer wearables have been validated for the full range of biometric outcomes they measure, including VO₂ max [18]. In fact, just one study evaluating the accuracy of Apple Watch for VO₂ max estimation has been published to date [32].

Evaluating the accuracy of VO₂ max estimates is fundamental to integrating wearable devices into individual health monitoring and research contexts [30]. For consumers, the reliability and validity of these measurements can inform decisions about personal health and fitness. For researchers, wearables hold the potential to revolutionise public health surveillance by enabling the collection of large-scale, population-level data without the logistical and financial constraints of laboratory-based methods [33]. On the contrary, inaccurate or inconsistent data may compromise the validity of such research, particularly when used to inform health policy or population health interventions. Validation is required to seal the fissure between technological innovation and evidence-based application in health and fitness settings.

The aim of this study was to evaluate the accuracy of VO₂ max estimates generated by Apple Watch in comparison to the gold standard method, indirect calorimetry. Specifically, the objectives were to assess the level of agreement between the two methods, and to quantify the degree of error in VO₂ max estimates from Apple Watch.

Materials and Methods

Results

Baseline characteristics

A total of 30 individuals participated in this study (mean age [SD], 31.86 [13.99] years; 50% female). Of the 30 participants, two (one male, one female) did not achieve the required criteria for VO₂ max, and as such 28 participants were included in the analyses. One of the excluded participants met only the threshold for rate of perceived exertion whereas the other failed to attain any of the criteria. Both excluded participants were wearing Apple Watch Series 9. The cardiorespiratory fitness level of each participant, determined by indirect calorimetry testing, was classified in accordance with the reference standard from the Fitness Registry and the Importance of Exercise National Database (FRIEND) [39]. 21 participants were classified as having either ‘Superior’ or ‘Excellent’ cardiorespiratory fitness, while nine individuals were classified as ‘Good’ or ‘Fair’. Participant baseline characteristics, including BMI and Fitzpatrick skin tone, are listed in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of the participants at baseline.

https://doi.org/10.1371/journal.pone.0323741.t001

Apple Watch agreement with the criterion

Overall, Apple Watch underestimated VO₂ max in comparison to indirect calorimetry. The mean difference was 6.07 mL/kg/min (SD 6.22; 95% confidence interval [CI] 3.77–8.38). Bland-Altman limits of agreement (LoA) indicated variability between the two measurement methods (lower LoA -6.11 mL/kg/min; upper LoA 18.26 mL/kg/min). The Bland-Altman plot is presented in Fig 1. The mean absolute percentage error (MAPE) was 13.31% (95% CI 10.01–16.61); mean absolute error (MAE) was 6.92 mL/kg/min (95% CI 4.89–8.94). To assess statistical power and effect size, a one‐sample t‐test was conducted comparing Apple Watch and criterion VO₂ max values. This analysis yielded a t‐value of –5.07 (df = 27, p < 0.001), indicating a significant difference with a large effect size (Cohen’s d = –0.96). A post-hoc power analysis (α = 0.05) confirmed the sample size (n = 28) provided >99% power, suggesting the study was sufficiently powered to detect a true difference between the measurement methods, thereby minimising the risk of Type II error. No trends in accuracy could be observed based on Apple Watch model. Results are summarised in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Statistical measures of agreement between Apple Watch and indirect calorimetry.

https://doi.org/10.1371/journal.pone.0323741.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Bland-Altman plot demonstrating the agreement between VO2 max measurements from Apple Watch and the criterion measure.

Upper LoA = upper limit of agreement; Lower LoA = lower limit of agreement.

https://doi.org/10.1371/journal.pone.0323741.g001

Discussion

This validation study evaluated the accuracy of VO₂ max estimates from Apple Watch compared to the gold-standard method of indirect calorimetry. Overall, Apple Watch underestimated VO₂ max, with a mean difference of 6.07 mL/kg/min (SD 6.22) and a MAPE of 13.31%. Bland-Altman limits of agreement indicated variability between the two measurement methods (lower -6.11 mL/kg/min; upper 18.26 mL/kg/min).

These findings align with those of a prior study evaluating Apple Watch Series 7, which reported a similar mean underestimation of 4.5 mL/kg/min [32]. However, important methodological differences exist. In the previous study, criterion VO₂ max values were obtained through a graded exercise test conducted on a cycle ergometer, rather than using a treadmill-based test. Apple Watch derives VO₂ max from walking, running or hiking activities [28], and comparison of VO₂ max measured through different exercise modalities – such as cycling compared with running – is not advised due to the inherent discrepancies in results they produce [40]. Typically, treadmill tests yield higher VO₂ max values, although the difference is smaller in well-trained cyclists [41]. This difference underpins the INTERLIVE consortium’s recommendation that criterion testing methods should closely align with the activity used by the wearable device to generate VO₂ max predictions [17].

While these methodological differences are important to consider, the authors noted distinct trends based on the fitness levels of participants [32]. For individuals with good or excellent fitness, Apple Watch demonstrated a propensity to underestimate VO₂ max, whereas among those with poor fitness, there was a tendency to overestimate. Although these authors found an overall underestimation, a separate meta-analysis of fourteen studies, which included multiple wearable device brands, reported a minor overestimation (pooled bias -0.09 mL/kg/min) [17]. Despite a very low bias, the limits of agreement were comparatively wide (lower -9.92 mL/kg/min; upper 9.74 mL/kg/min), suggesting both positive and negative differences in estimates of a relatively large magnitude, which collectively yielded a bias close to zero. A propensity to both under- and overestimate may arise from prediction algorithms that are developed using population-level data or mean values for specific demographics. Given our study sample predominantly consisted of individuals with high cardiorespiratory fitness levels, this may constitute one explanation for our finding of underestimation. Importantly, it highlights the need to consider participant demographic characteristics, as well as multiple statistical measures of agreement, when interpreting results.

Our findings, combined with the broader literature, underscore the challenge of deriving maximal oxygen consumption from submaximal exercise for wearable devices. Like conventional prediction equations, Apple’s prediction algorithm is founded on heart rate response to exercise [28]. The inclusion of machine learning techniques and additional motion sensor measurements is also probable, although the precise nature of the algorithm remains undisclosed [30]. In clinical settings, submaximal exercise testing is often used alongside prediction equations to estimate VO₂ max, owing to safety and feasibility concerns [35]. These equations predict VO₂ max based on an individual’s physiological response to exercise of a quantified workload, and six assumptions – outlined by the ACSM – must be achieved to ensure optimal accuracy [35]. They include a minimal difference between actual and predicted maximal heart rate, knowledge of medication or substances affecting heart rate, and the existence of a linear relationship between heart rate and work rate. Yet, accuracy from traditional prediction equations can vary significantly [42–44]. A study (n = 541) evaluating three different submaximal exercise tests for estimating maximal oxygen uptake reported a standard error of the estimation ranging from 3.7 to 4.5 mL/kg/min [45]. In a separate study examining the Chester Step Test, the predicted aerobic capacity had a standard error of 3.9 mL/kg/min [19,46]. Comparatively, measurement error for laboratory-based indirect calorimetry has been estimated at ±5%, with a meta-analysis of 39 studies reporting a mean standard error of 2.58 mL/kg/min [40,47]. Considering this literature, the results of our validation study indicate that Apple Watch VO₂ max predictions fall approximately two to three standard deviations beyond the typical error of the gold standard measurement. However, they align more closely with the magnitude of error reported for traditional VO₂ max predictions derived from submaximal exercise.

While machine learning and continuous monitoring are distinct merits of Apple Watch compared to traditional prediction methods, the uncontrolled nature of the exercise used to generate Apple Watch VO₂ max estimates presents a considerable challenge. Limited information of external factors that affect heart rate response to exercise – including exercise surface, environmental temperature, or alcohol intake – can adversely impact predictions [28]. Moreover, heart rate response to lower-intensity exercise – often used to estimate VO₂ max – varies considerably between individuals [48]. Estimates based on such exercise may misinform predictions. Due to the proprietary nature of Apple’s algorithm, we remain unaware of the specific data, or features, used to train and develop the prediction model. Therefore, we cannot be certain what factors most influence accuracy, or in which instances predictions are most reliable [30]. Considering this, a defined testing protocol – requiring users to input specific information and conduct a particular type of exercise – may reduce the level of uncertainty and error. Despite the imprecision of predictions from submaximal exercise – both from traditional testing methods and wearable devices – its widespread clinical use illustrates its value. The potential for Apple Watch to provide a commensurate assessment method that can be conducted independently by members of the general public warrants careful consideration. To investigate this, validation of contemporary technology is apt.

Continual, or ‘living’, validation is essential to keep pace with the iterative commercial ecosystem. The annual update cycle of Apple Watch poses a significant challenge to evaluating current technology, and the prolonged academic publication process often means that the hardware or software under validation has been discontinued by the time of publication [18,30]. This is accentuated by the rapid advancement of machine learning – and the associated updates to watchOS – which play an increasingly important role in providing accurate measurements due to the complexity of interpreting multiple sensor measurements and sensitive photoplethysmography waveforms [49,50]. Amidst these challenges, this study provides a timely evaluation of Apple’s most recent optical heart rate sensor and its latest VO₂ max prediction algorithm [28,51]. It is the first study to evaluate Apple Watch Series 9 and Ultra 2. As wearable technology evolves, agile validation will be critical to realising the potential of consumer wearables in large-scale public health applications and personalised fitness monitoring.

Central to this potential is the self-directed nature of wearable device measurements, and this study’s methodology was designed to reflect this. Participants were not required to adhere to a prescribed protocol to generate a VO₂ max prediction. Rather, they were solely informed of the requirements, meaning that estimates were generated through activities that more accurately reflected each individual’s real-world use, enhancing the ecological validity of our findings. Additionally, a robust statistical approach was used, guided by recommendations of the INTERLIVE expert consortium [17], and a sex-balanced study sample was recruited. However, the predominance of participants with high levels of cardiorespiratory fitness represents the study’s most significant limitation. Only one participant was classified as having ‘Fair’ cardiorespiratory fitness, while all others were categorised as ‘Good’, ‘Excellent’, or ‘Superior’. This limits the generalisability of our findings to populations with lower fitness levels, including clinical cohorts. Additionally, the sample primarily consisted of younger individuals, despite efforts to recruit participants across a broad age range. Another limitation is the uncontrolled procedure of generating Apple Watch VO₂ max estimates. While this protocol reflects real-world use, it introduces unquantified external variables that may have influenced prediction accuracy. Lastly, only one VO₂ max estimate was collected per participant, preventing analysis of intra-subject variability, which may have provided insights into reliability and changes in measurement accuracy over time.

The findings of this study have important implications for the use of wearable devices such as Apple Watch in clinical and research contexts. VO₂ max estimates from Apple Watch offer a cost-effective and scalable alternative to laboratory-based testing, enabling studies involving diverse populations, as well as digital endpoints for clinical trials. However, given the observed variability and underestimation of VO₂ max in this study, caution is warranted when interpreting the data in clinical or research settings. While Apple Watch may provide a suitable estimation of aerobic capacity for general fitness monitoring, our findings suggest that estimates are not sufficiently accurate to inform clinical decision-making. Future research should include individuals with diverse cardiorespiratory fitness levels to enhance generalisability, support the refinement of predictive algorithms, and develop frameworks for ongoing evaluation of wearable devices. Such efforts are essential to bridging the gap between technological innovation and evidence-based practice, ultimately enhancing the utility of wearable devices for individual health monitoring and public health applications.

Conclusions

This study found that Apple Watch underestimated VO₂ max compared to indirect calorimetry in a sample predominantly composed of individuals with high cardiorespiratory fitness levels. The margin of error more closely aligned with that of VO₂ max derived from conventional submaximal exercise testing than with VO₂ max obtained via indirect calorimetry. The unstructured nature of the exercise used to generate Apple Watch VO₂ max predictions introduces external factors – such as environmental conditions and variations in heart rate response to exercise – that are not fully accounted for by the device. Nevertheless, Apple Watch holds promise as a practical and accessible alternative to conventional submaximal exercise testing. Its potential clinical utility warrants further investigation, and ongoing validation is required to evaluate its accuracy as hardware and software evolve. Future research should focus on expanding validation efforts to include diverse populations, ensuring these devices can be effectively integrated into both individual health monitoring and broader public health applications.