Prediction of the oxygen uptake patterns during an incremental exercise test using long short - term memory in electromyography

Abstract

Prediction models of the oxygen uptake (VO2) from electromyograms (EMG) of the lower limb and respiratory muscles during an incremental exercise test were examined. Healthy male adults (n=15) underwent an incremental exercise test using a cycle ergometer. To predict the patterns of VO2, we used a type of recurrent neural network, the long short-term memory. The measured patterns of VO2 were used as training data for deep learning, and two prediction models as input values were set: a lower limb muscle model and a respiratory muscle model. In the lower limb muscle model, EMGs of the rectus femoris and vastus lateralis were input. In the respiratory muscle model, EMGs of the sternocleidomastoid and inspiratory time were input. The patterns of both the measured VO2 and predicted VO2 increased during the exercise test. The histogram showed a peak difference between the measured and predicted VO2 of 0 and 0.5 mL/kg/min. The Bland-Altman plots for both models demonstrated that most of the data were distributed within the range of agreement. The root mean square error (RMSE) during the exercise period was 2.1 ± 0.7 mL/kg/min for the lower limb muscle model and 2.8 ± 1.1 mL/kg/min for the respiratory muscle model, and the RMSE increased with the increasing course of time. In the cycle ergometer task, each model enable the estimation of the pattern of the VO2. Mild to moderate exercise intensity was suitable for the prediction of VO2 patterns by electromyography.