EMG feature extraction and muscle selection for continuous upper limb movement regression

Achieving a versatile control of exoskeletons or prostheses requires accurate, robust and fast predictions of upcoming human movements. Electromyographic (EMG) signals are increasingly being used for this purpose as they theoretically allow to estimate joint torques in advance due to the electromechanical delay. However, their performance in continuously predicting joint torques strongly depends on the location of the sensors and on the feature extracted from the signal, whose impact on the accuracy is little-known in this context. In the present paper, the influence of different EMG features and muscles on the performance to predict both multi- and single-joint torques have been analyzed, while moving in a parasagittal plane hindered by a viscous force field applied by an exoskeleton. The existence of an accuracy–latency tradeoff is first highlighted, whether with frequency or time-domain features, the latter providing the best tradeoff. Second, the impact of muscle selection on the quality of the joint torques prediction is illustrated. In particular, the anterior deltoid, the long triceps, the brachialis, and the brachioradialis muscles are shown to be the most important to record during movements in a parasagittal plane. The present results pave the path towards the design of reactive and accurate exoskeletons and prostheses controllers based on EMG signals • Feature extraction shows a trade-off between latency and accuracy. • Time-domain features consistently outperform frequency-domain features. • In time-domain, window length impacts performance more than the feature itself. • Increasing muscle count boosts accuracy but with diminishing returns. • Analyzing certain muscles rather than others is more effective for accurate torque prediction.