Watch VO2max estimates are close to lab values on average but can be roughly 10 ml/kg/min out for an individual, and watch-style race predictions tend to be optimistic
Interpretation
A meta-analysis of 14 validation studies (Molina-Garcia 2022) found near-zero average bias for exercise-based estimates (-0.09 ml/kg/min) but limits of agreement of about ±10, and overestimation by resting-based devices (2.17; limits -13.07 to 17.41). Later single-device studies agree on individual error: MAPE 7.05% in 19 adults (Carrier 2025), 13.31% with 6.07 underestimation in 30 adults (Lambe 2025), 7.2-7.9% in 35 athletes with larger underestimation (6.3) in the highly trained (Engel 2026), and overestimation of 2.6-2.9 units in 20 runners (Howard 2026). Race predictions: half-marathon predicted 129.97 vs actual 137.55 min (Howard 2026); the Riegel formula was at least 10 min too fast for the marathon for half of 2,303 runners (Vickers 2016). Graded moderate: consistent direction, but small samples, few women reported, no triathletes, and device software changes quickly.
Practical implication
Read watch VO2max as a trend over weeks, check it against field tests and race results, and don't set race pace from watch predictions, which assume a standalone run on fresh legs. Triathlon runs come after a swim and bike, so predictions are an upper limit at best; this transfer is inference, as no study has tested watch predictions for triathlon.
Sources
6 sources, independently verified.
Wearables using resting information overestimated VO2max (bias 2.17 ml/kg/min; limits of agreement -13.07 to 17.41); exercise-based algorithms showed lower error (bias -0.09; limits -9.92 to 9.74). Exercise-based estimates may suit population-level use, but individual error is large. Proposes six-domain validation standards.
A multisport watch (2019 model) estimated VO2max with MAPE 7.05% and Lin's concordance 0.73 against 30-s averaged laboratory data, meeting the authors' criteria (MAPE < 10%, CCC > 0.7); its pulse-oximetry readings failed validation.
After 5-10 days of wear, a general-purpose smartwatch (2023 models) underestimated treadmill-measured VO2max by 6.07 ml/kg/min (95% CI 3.77-8.38), limits of agreement -6.11 to 18.26, MAPE 13.31%, mean absolute error 6.92 ml/kg/min.
A running watch (2019 model) underestimated VO2max after two 15-min outdoor runs by 4.73 and 4.05 ml/kg/min (MAPE 7.9% and 7.2%; ICC 0.71 and 0.75). Moderately trained athletes: MAPE 2.8-4.1%; highly trained (lab VO2max > 59.8): underestimated by 6.3 ml/kg/min, MAPE 9.4-10.4%. Published online 2025.
A running watch (2023 model) overestimated VO2max at baseline (43.95 vs 41.10 ml/kg/min) and after 12 weeks (45.00 vs 42.43); its half-marathon predictions (129.97 min) were significantly faster than actual race times (137.55 min). MAPE for VO2max and prediction was under 10%. Online ahead of print; no volume or issue yet.
The Riegel race-time formula was well calibrated for races up to the half-marathon but dramatically underestimated marathon time, giving times at least 10 min too fast for half of runners; models including training data (e.g. weekly mileage) predicted marathon time better (mean squared error 381 for Riegel vs 228 and 208). Self-reported survey data.