Hand-movement Prediction from EMG with LSTM-Recurrent Neural Networks

In this paper we present an approach based on shallow recurrent long short-term memory neural networks for the prediction of hand kinematics for hand-prosthesis control from data acquired via high-density surface electromyography (HD-sEMG). We used 134-channel HD-sEMG recordings from seven participants while performing multiple repetitions of 13 hand movements. A CyberGlove II was used to simultaneously record 18 degrees of freedom (joint angles) used as ground truth for predicting the hand movements. Traditional features were calculated over 100 ms windows and fed to the network. Specifically we used: Mean Absolute Value (MAV), variance, and number of zero-crossings. Our results indicate that: (a) a small number of channels is sufficient to make accurate predictions, (b) many features are redundant, and MAV is sufficient for the job, (c) the simple neural network architecture we propose is effective in this task. These findings have important implications in terms of computational efficiency and memory storage, which are important considerations in relation to implementability in the typically very low-power and low-resources computers onboard of hand prostheses.