Decoding finger movement patterns from microscopic neural drive information based on deep learning.

Recent development of surface electromyogram (sEMG) decomposition technique provides a good basis of decoding movements from individual motor unit (MU) activities that directly representing microscopic neural drives. How to interpret the function and contribution of each decomposed MU to macroscopic movements remains unclear. The objective of this study is to decode finger movement patterns by establishing a relationship between individual MU activities and movements. In this study, high-density sEMG (HD-sEMG) data were recorded by a 16 × 8 electrode array from finger extensor muscles of 10 subjects performing 10 finger movement patterns. The progressive FastICA peel-off algorithm was first applied to decompose the HD-sEMG data to obtain microscopic neural drives in terms of MU firing sequences and their corresponding action potential waveforms. Then, convolutional neural network was used for classification of the decomposed MUs by characterizing their spatial waveforms spanned over all channels of the array. On this basis, a fuzzy weighted decision strategy was designed to give a final decision of movement pattern recognition, where function of an individual MU was measured in the form of contributing into all movement patterns with different weights to solve the issue of MUs shared among multiple patterns due to muscle co-activation. The proposed method yielded an average accuracy approximating to 100%, and it outperformed other common MU-based methods or conventional myoelectric classification methods using macroscopic sEMG features (p <  0.05). The proposed method has a wide application prospect in the field of human-machine interaction and precise motor control.