Performance and limitations of smartwatch estimation of VO2max during free-living activity in non-athletes

Background. Maximal oxygen consumption (VO ₂ max) is the gold standard for assessing cardiorespiratory fitness (CRF), and it is a strong predictor of clinical outcomes, but it is measured during cardiopulmonary exercise testing (CPET), which is expensive and time-consuming. Smartwatches provide a great opportunity to assess CRF at scale. Although smartwatches have been previously validated in young and physically active individuals, their accuracy in sedentary individuals remains undetermined and this limits the use smartwatch VO ₂ max in large clinical or epidemiological studies. This study aims to determine the accuracy of smartwatch VO ₂ max in a large group of individuals with varying underlying CRF. Methods. N=203 adults (160 (79%) female; median (interquartile range; IQR) 55 (32,61) years old) from the CONVALESCENCE case-control study of long COVID, performed a maximal CPET on a semi-recumbent ergometer and wore a consumer-grade smartwatch for three months. Smartwatch-estimated VO ₂ max was compared to reference VO ₂ max measured by CPET and predicted VO ₂ max calculated using standard non-exercise equations. Agreement was assessed using Bland-Altman analysis (Bias [Limits of Agreement; LoA]), error (ΔVO ₂ max), absolute error (|ΔVO ₂ max|), and Pearson’s correlation coefficient (r). Results. CPET-measured VO ₂ max was 22.2 (17.5, 27.3) ml/kg/min. Smartwatch-estimated VO ₂ max correlated moderately with CPET-measured VO ₂ max (r=0.65) and strongly with VO ₂ max predicted using non-exercise equations (r=0.92). Bland-Altman plots showed a large positive bias and wide LoA (14.5 [2.2, 26.8 ml/kg/min]). Inaccuracy of smartwatch-estimated VO ₂ max (ΔVO ₂ max) linearly increased with increasing mismatch between expected and measured CRF (correlation with %pVO ₂ max r=-0.81). The absolute error was larger in those with lower-than-expected CRF (%pVO ₂ max<100%) and in those engaging in light intensity activities, while no differences were observed when comparing participants with or without long COVID or symptoms of fatigue. Multiple linear regression analysis of the absolute error identified the mismatch between expected and measured CRF (%pVO ₂ max) as the main contributor to smartwatch inaccuracy. Conclusions. In non-athletes, smartwatch-derived VO ₂ max correlated with reference values but did not provide additional accuracy to VO ₂ max prediction based on sex, age, weight and height. Inaccuracy significantly increased in those with lower-than-expected CRF. These findings highlight current limitations in the use of smartwatch VO ₂ max for clinical applications and large epidemiological studies. Trial registration number: Z6364106/2021/06/04.