Prediction of oxygen uptake in dragon boat athletes using machine learning and multimodal physiological signals

Objective This study developed and evaluated a machine learning model for predicting oxygen uptake (VO 2 ) in dragon boat athletes using wearable signals and compared performance across signal combinations and device groups. Methods Twelve male collegiate dragon boat athletes completed nine tests grouped into six task types (T1–T6): 200-, 500-, and 1000-m all-out paddling tests; an incremental test; a five-stage constant-load test; and four supramaximal-intensity tests pooled as T6. Heart rate, respiratory rate, device-derived minute ventilation, and muscle-oxygenation signals were collected. Fifteen signal combinations (G1–G15) and six device groups (A–F) were evaluated using a multilayer perceptron model. Performance was assessed using mean absolute error (MAE) and root mean square error (RMSE), with mean absolute percentage error (MAPE) also reported. Results G13–G15 showed comparable performance, with no significant differences. G14 yielded the lowest MAE and MAPE (2.21 mL·kg −1 ·min −1 and 9.48%), whereas G13 yielded the lowest RMSE (3.19 mL·kg −1 ·min −1 ). No significant differences were observed among signal combinations within Groups A, C, D, E, or F. Group F showed the lowest MAPE but did not differ significantly from Group B, whereas Group C showed significantly higher MAPE than Groups B, D, E, and F. Performance did not differ significantly across exercise tasks. During the 1000-m test, predicted and measured VO 2 were strongly correlated for G15 ( r = 0.88), with a mean prediction error of 0.44 mL·kg −1 ·min −1 and descriptive 95% limits of agreement from −6.32 to 7.20 mL·kg −1 ·min −1 . Conclusion Cardiorespiratory signals provided useful information for wearable-based VO 2 prediction, whereas the contribution of muscle-oxygenation signals depended on the input configuration. G13–G15 should be regarded as comparably performing configurations, and bilateral muscle-oxygenation monitoring was not consistently superior to unilateral monitoring. Multimodal wearable signals show potential for non-invasive VO 2 monitoring in dragon boat athletes, although larger samples and independent validation are required.