Prediction of oxygen uptake kinetics during heavy-intensity cycling exercise by machine-learning analysis.

Non-intrusive estimation of oxygen uptake (V̇O) is possible with wearable sensor technology and artificial intelligence. V̇O kinetics have been accurately predicted during moderate exercise using easy-to-obtain sensor inputs. However, V̇O prediction algorithms for higher intensity exercise with inherent nonlinearities are still being refined. The purpose of this investigation was to test if a machine learning model can accurately predict dynamic V̇O across exercise intensities, including slower V̇O kinetics normally observed during heavy- compared to moderate-intensity exercise. Fifteen young healthy adults (7 females; peak V̇O: 42±5 mL·min·kg) performed three different pseudorandom binary sequence (PRBS) exercise tests ranging in intensity from low-to-moderate, low-to-heavy, and ventilatory threshold-to-heavy work rates. A temporal convolutional network was trained to predict instantaneous V̇O, with model inputs including heart rate, percent heart rate reserve, estimated minute ventilation, breathing frequency, and work rate. Frequency domain analyses between V̇O and work rate were used to evaluate measured and predicted V̇O kinetics. Predicted V̇O had low bias (-0.017 L·min, 95% limits of agreement: [-0.289, 0.254]), and was very strongly correlated (=0.974, <0.001) with the measured V̇O. The extracted indicator of kinetics, mean normalized gain (MNG), was not different between predicted and measured V̇O responses (main effect: =0.374, η=0.01), and decreased with increasing exercise intensity (main effect: <0.001, η=0.64). Predicted and measured V̇O kinetics indicators were moderately correlated across repeated measurements (MNG: =0.680, <0.001). Therefore, the temporal convolutional network accurately predicted slower V̇O kinetics with increasing exercise intensity, enabling non-intrusive monitoring of cardiorespiratory dynamics across moderate- and heavy-exercise intensities.