Classification of Individual and Combined Finger Flexions Using Machine Learning Approaches

Abstract

As the number of amputees in the world steadily increases each year, the need for fully functional prosthesis is at an all-time high. Enabling a wide variety of possible movements with fast, accurate, and fluid control is one of the main goals for a successful electric prosthesis. The use of electromyography for the detection of the intended movements allows for increased practicality of the device and easy and natural control by the user. The implementation of a fast and cheap classification method for the measured signals allows for a significantly lower price of the proposed prosthesis, thus making it available to a wider margin of the population while also maintaining the expected quality of the product. In this paper we compare multiple different machine learning algorithms for the classification of individual and combined finger flexions from eight different participants using two-channel electromyography, in order to achieve an optimal generalized classifier that allows for accurate results while providing the least complex solution to the problem. By using only two EMG sensors we can achieve a much more practical and cheaper realization of an electric prosthesis, which is the main goal of this work. The best performer out of the tested algorithms is then additionally evaluated on the same movements when they are carried out by a single participant, thus simulating a real-world case where the prosthesis would be used by a single individual. Finally, we analyze classification errors, discuss the nature of their occurrences, and propose possible solutions and future improvements.