Synergy-Based Neural Interface for Human Gait Tracking with Deep Learning

Abstract

Neural information decomposed from electromyography (EMG) signals provides a new path of EMG-based human-machine interface. Instead of the motor unit decomposition-based method, this work presents a novel neural interface for human gait tracking based on muscle synergy, the high-level neural control information to collaborate muscle groups for performing movements. Three classical synergy extraction approaches include Principle Component Analysis (PCA), Factor Analysis (FA), and Nonnegative Matrix Factorization (NMF), are employed for muscle synergy extraction. A deep regression neural network based on the bidirectional gated recurrent unit (BGRU) is used to extract temporal information from the synergy matrix to estimate joint angles of the lower limb. Eight subjects participated in the experiment while walking at four types of speed: 0.5km/h, 1.0km/h, 2.0km/h, and 3.0km/h. Two machine learning methods based on linear regression (LR) and multilayer perceptron (MLP) are set as the contrast group. The result shows that the synergy-based approach's performance outperforms two contrast methods with ${R}_{\textit {var}}^{2}$ scores of 0.830.88. PCA reaches the highest performance of 0.871±0.029, corresponding to RMSE of 3.836°, 6.278°, 2.197° for hip, knee, and ankle, respectively. The effect of walking speed, synergy number, and joint location will be analyzed. The performance shows that muscle synergy has a good correlation will joint angles which can be unearthed by deep learning. The proposed method explores a new way for gait analysis and contributes to building a novel neural interface with muscle synergy and deep learning.