Trunk compensation electromyography features purification and classification model using generative adversarial network

Abstract

Reducing trunk compensation can improve the effect of rehabilitation training. Therefore, automatic detection of patient compensation can prompt patients to adjust their exercise mode. However, the classification accuracy of trunk compensation in stroke patients is not satisfactory due to individual movement variability. In this paper, we introduced a novel method for detecting the trunk compensatory patterns of stroke patients based on GAN. Our strategy relies on a three-stage approach to realize the trunk compensatory patterns detection：(i) The bandpass filter and sample entropy algorithm are used for sEMG data noise reduction and motion segment division; (ii) Extract the time-domain features of sEMG and; (iii) Features purified and patterns classified by GAN-KNN. Ten stroke patients were recruited to participate in the experiments to verify the feasibility of this method. All stroke participants performed three reaching rehabilitation training tasks using the unaffected and affected arm, while the sEMG of nine trunk muscles were recorded. Results show that this method achieved state-of-the-art detection performance (accuracy = 94.58 ± 1.15 %) of trunk compensatory patterns for stroke patients, which was significantly higher than the detection performance based on SVM classifier (average accuracy = 83.54 ± 0.52 %; macro-F1 = 0.84) and KNN classifier (average accuracy = 74.79 ± 1.42 %; macro-F1 = 0.75). The result demonstrated the feasibility and effectiveness of this method. This study has the potential to assist stroke patients to autonomously correct trunk compensation or to apply to the human body status recognition of rehabilitation assistant robot.