EMG-driven hand model based on the classification of individual finger movements. (April 2020)

• Record Type: Journal Article
• Authors: Arteaga, Maria V.; Castiblanco, Jenny C.; Mondragon, Ivan F.; Colorado, Julian D.; Alvarado-Rojas, Catalina
• Is Part Of: Biomedical signal processing and control. Volume 58(2020)
• Publication Date: 2020-04
• Subjects: Electromyography -- Signal processing algorithms -- Machine learning -- Hand model -- Inverse and forward dynamics; 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.2019.101834
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We have applied machine learning to identify 6 finger gestures from EMG signals. Time/frequency features, ANN, SVM and k-NN algorithms were used for classification. Fine and Weighted k-NN showed a better performance with accuracy of 97%. The gestures were turned into a joint trajectory using interpolation methods. We then reconstruct the finger kinematics and simulate the dynamics. Experiments were carried out to create an EMG database from 20 control subjects. The correlation between the joint trajectories and the tracked hand-motion was 0.91. Abstract: The recovery of hand motion is one of the most challenging aspects in stroke rehabilitation. This paper presents an initial approach to robot-assisted hand-motion therapies. Our goal was twofold: firstly, we have applied machine learning methods to identify and characterize finger motion patterns from healthy individuals. To this purpose, Electromyographic (EMG) signals have been acquired from flexor and extensor muscles in the forearm using surface electrodes. Time and frequency features were used as inputs to machine learning algorithms for recognition of six hand gestures. In particular, we compared the performance of Artificial Neural Networks (ANN), Support Vector Machines (SVM) and k-Nearest Neighbor (k-NN) algorithms for classification. Secondly, each identified gesture was turned into a joint reference trajectory by applying interpolation methods. This allowed us to reconstruct the hand/finger motion kinematics and to simulate the dynamics of each motion pattern. Experiments were carried out to create an EMG database from 20 control subjects, and a VICON camera tracking system was used to validate the accuracy of the proposed system. The average correlation between the EMG-based generated joint trajectories and the tracked hand-motion was 0.91. Furthermore, statistical analysis applied to 14 different SVM, ANN and k-NN configurations showed that Fine k-NN and Weighted k-NN have a better performance for the classification of gestures (98% of accuracy). In a future, the trajectories controlled by EMG signals could be applied to an exoskeleton or hand-robotic prosthesis for rehabilitation.