Figure 3.

EMG signal processing, decomposition and dimensionality reduction. For each grip type (top), the central 10 s of EMG recordings was concatenated to obtain 70-s long EMG signals (10 s per grip type). These concatenated monopolar signals were either differentiated (selection of one bipolar channel per muscle; see Fig. 4A) and rectified or decomposed with the CKC algorithm into motor neuron spike trains. Both rectification and CKC are nonlinear operations. Additionally, CKC reduces the dimensionality of the original signals. The results of these two nonlinear operations were low-pass filtered to generate smooth myoelectric signals with a bandwidth of 2.5 Hz. The two smoothed signals were factorized using NMF to extract muscle synergies (where one channel is related to a muscle) and motor neuron synergies, respectively. NMF is a linear dimensionality reduction operation.