Estimation of continuous thumb angle and force using electromyogram classification

• Published in: International journal of advanced robotic systems 2016-09, Vol.13 (5)
• Main Authors: Siddiqi, Abdul Rahman, Sidek, Shahrul Naim
• Format: Article
• Language: English
• Subjects: Accuracy; Classification; Classifiers; Data mining; Feature extraction; Fingers & toes; Muscles; Neural networks; Prostheses; Signal classification; Signal processing; Studies; Support vector machines
• ISSN: 1729-8806; 1729-8814
• DOI: 10.1177/1729881416658179

Human hand functions range from precise minute handling to heavy and robust movements. Remarkably, 50% of all hand functions are made possible by the thumb. Therefore, developing an artificial thumb that can mimic the actions of a real thumb precisely is a major achievement. Despite many efforts dedicated to this area of research, control of artificial thumb movements in resemblance to our natural movement still poses as a challenge. Most of the development in this area is based on discontinuous thumb position control, which makes it possible to recreate several of the most important functions of the thumb but does not result in total imitation. This work looks into the classification of electromyogram signals from thumb muscles for the prediction of thumb angle and force during flexion motion. For this purpose, an experimental setup is developed to measure the thumb angle and force throughout the range of flexion and simultaneously gather the electromyogram signals. Further, various features are extracted from these signals for classification and the most suitable feature set is determined and applied to different classifiers. A “piecewise discretization” approach is used for continuous angle prediction. Breaking away from previous research studies, the frequency-domain features performed better than the time-domain features, with the best feature combination turning out to be median frequency–mean frequency–mean power. As for the classifiers, the support vector machine proved to be the most accurate classifier giving about 70% accuracy for both angle and force classification and close to 50% for joint angle–force classification.