Robust EMG Pattern Recognition with Electrode Donning/Doffing and Multiple Confounding Factors

Traditional electromyography (EMG) pattern recognition did not take into account confounding factors such as electrode shifting, force variation, limb posture, etc., which lead to a great gap between academic research and clinical practice. In this paper, we investigated the robustness of EMG pattern recognition under conditions of electrode shifting, force varying, limb posture changing, and dominant/non-dominant hand switching. In feature extraction, we proposed a method for threshold optimization based on Particle Swarm Optimization (PSO). Compared with the traditional trail & error method, it can largely increase the classification accuracy (CA) by 10.2%. In addition, the hybrid features integrated with discrete Fourier transform (DFT), wavelet transform (WT), and wavelet packet transform (WPT) were proposed, which increased the CA by 30.5%, 25.4%, 22.9%, respectively. We introduced probabilistic neural network (PNN) as a new classifier for EMG pattern recognition, and reported the CA's obtained by a large variety of features and classifiers. The results showed that the combination of DFT_MAV2 (a novel feature based on DFT) and PNN reached the best CA (45.5%, 14 motions, validated on different hands without re-training).