A Spatiotemporal Deep-Learning Model for Force Estimation From Surface Electromyography

The noninvasive estimation of muscle force from muscular activations is of particular interest for different clinical applications, such as prosthesis control and neurorehabilitation. Surface electromyography (sEMG) enables to measure the electrical activity of a muscle and has been shown in previous works to be associated with force production. However, a general model mapping sEMG to force in real-life situations is yet to be established, hampering the clinical translation of sEMG-based force estimation methods. In this study, we aim at estimating force from sEMG during dynamically changing force-levels achieved with isometric muscular contractions. 64 High Density sEMG (HD-sEMG) channels were acquired from the biceps brachii of 50 healthy subjects in order to record both spatial and temporal distribution of the muscle's electrical activity. The participants performed isometric contractions ranging between 0 and 80% of their maximum voluntary contraction. A deep-learning strategy was adopted in this work. The normalized bipolar HD-sEMG signals were used as input to an adapted three-dimensional version of the temporal convolution network (TCN) which jointly extracts spatiotemporal features. The obtained result supports that force can be estimated with a normalized root mean squared error of 29.2 ± 13.1% and an R2 value of 64.9 ± 25.7%. This shows that deep-learning holds promise for noninvasive estimation of varying force from sEMG signals.