Motion intention prediction of upper limb in stroke survivors using sEMG signal and attention mechanism. (September 2022)

• Record Type: Journal Article
• Authors: Li, Juncheng; Liang, Tao; Zeng, Ziniu; Xu, Pengpeng; Chen, Yan; Guo, Zhaoqi; Liang, Zhenhong; Xie, Longhan
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Motion intention recognition -- sEMG -- Attention mechanism -- Deep learning -- Rehabilitation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103981
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We further explain the used data sets and the principle and function of T-SNE. We further explain the method to solve the over fitting issue. We retested the effectiveness of the model with a separate independent test set. We calculate the time complexity of the model and examine the effectiveness of the proposed model. We added a comparison analysis of the three models in the discussion section. Abstract: The upper limb movement of stroke survivors has strong specificity and involuntary activation of muscles and other non-ideal factors. The prediction method suitable for healthy people often declines accuracy when applied to stroke survivors. The precise perception of the patient's motion intention is helpful for the patient to use the rehabilitation robot for rehabilitation training. Current research focuses on data acquisition, preprocessing, feature extraction, and classifier selection. Some researchers have proposed effective methods, but they have disadvantages such as complexity, high cost, and low generalization. In this paper, we proposed a new solution to the problem of significant interference of patients' sEMG data: (i) Embedding the attention mechanism into the deep residual network so that the attention module can entirely focus on the key features to improve the network's learning ability of features. (ii) The soft thresholding module is embedded into the deep residual network as a building unit, and the threshold is automatically set to eliminate the interfering noise. We designed an experiment to acquire sEMG signals from eight muscles of ten patients during six preset movements and adopted a 10-fold cross-validation method to verify the feasibility of the proposed method. The length of the data processing window, the prediction accuracy of different movements, and various models' classification effect are compared. The results show that compared with ResNet (average accuracy = 84.94 %) and CNN (average accuracy = 78.47 %), the proposed method has higher classification accuracy, with an average accuracy of 93.11 %, which proves the feasibility of the proposed method. This study can be applied to improve the efficiency of rehabilitation training for stroke survivors.