Fig. 5: Results on the high-density EMG decomposition.

a Decomposition of experimental high-density surface electromyography (HD-sEMG) recordings into underlying spinal motor neuron activities. The results obtained with the neural network (NN) were tested against the decomposition by a reference blind source separation method and manual editing by an expert operator. MU stands for motor unit. b Myoelectric digital twins were used to generate motor unit action potential (MUAP) templates for different muscles and different model parameters (tissue conductivities, fibre properties, tendon sizes, etc). 64 sets, each containing 5 simulated MUAPs for 130 electrodes, were used for pre-training. c Rate of agreement (%) between the neural network MU activity predictions and the decomposition algorithm on one second of wrist flexor HD-sEMG signal. Data is presented as median values over 39 motor units from nine participants +/− the interquartile range on the bounds of the box, with the lowest and highest values as the whiskers. Both outputs were converted to timestamps using a two class K-means clustering. The neural network using a gated recurrent unit (GRU) network that was pre-trained using simulated EMG signal significantly outperformed a GRU with random initialisation (two-tailed Wilcoxon signed-rank test, Z = 4.0, p = 0.00006, median difference 8.1 using Hodges-Lehmann estimator, 95% CI 3.4 to 13.3 using method of Walsh averages).