A long short-term memory modeling-based compensation method for muscle synergy

Muscle synergy containing temporal and spatial patterns of muscle activity has been frequently used in prediction of kinematic characteristics. However, there is often some discrepancy between the predicted results based on muscle synergy and the actual movement performance. This study aims to propose a new method for compensating muscle synergy that allows the compensated synergy signal to predict kinematic characteristics more accurately. The study used the change of direction in running as background. Non-negative matrix factorisation was used to extract the muscle synergy during the change of direction at different angles. A non-linear association between synergy and the height of pelvic mass centre was established using long and short-term memory neural networks. Based on this model, the height fluctuations of the pelvic centre of mass are used as input and predict the fluctuations of the synergy which were used to compensate for the original synergy in different ways. The accuracy of the synergies compensated in different ways in predicting pelvic centre of mass movement was then assessed by back propagation neural networks. It was found that the compensated synergy significantly improves accuracy in predicting pelvic centre of mass displacement (R², p < 0.05). The predicted results of all-compensation are significantly different from actual performance in the end-swing (p < 0.05). The predicted results of half-compensation do not differ significantly from the actual performance, and its damage to the original synergy is smaller and does not increase with angle compared to all-compensation. The all-compensation may be affected by individual variability and lead to increased errors. The half-compensation can improve the predictive accuracy of the synergy while reducing the adjustment to the original synergy.