74 IEEE TRANSACTIONS ON HUMAN-MACHINE SYSTEMS, VOL. 45, NO. 1, FEBRUARY 2015 Hand and Wrist Movement Control of Myoelectric Prosthesis Based on Synergy Jiaxin Ma, Nitish V. Thakor, Fellow, IEEE, and Fumitoshi Matsuno, Member, IEEE Abstract—This study proposes a method to control a prosthetic hand by EMG signals based on muscle synergies. The muscle syn- ergy model suggests a framework to transform commands of the central nervous system to a set of complex muscular movements. Using this method, we have tried to realize the proportional con- trol of multiple degrees of freedom (DOF). This study focuses on controlling four kinds of hand/wrist movements of the prosthesis: open, close, pronate, and supinate. The nonnegative matrix factor- ization (NMF) algorithm is used to map muscle activities into these four movements through the calculation of a muscle synergy ma- trix. An EMG feature selection process along with a control scheme has been added, which smooths the output thereby stabilizing the movements. Ten healthy subjects performed an online experiment comprised of two tests: 1) proportional control on single DOF, and 2) simultaneous control of multiple DOFs. The results indicate that fluid hand/wrist movements could be estimated from EMG. The average R 2 values achieved by all subjects for the single-DOF test and the multiple-DOF test are 0.97 and 0.93, respectively. Index Terms—Degrees of freedom (DOF), electromyogram (EMG), nonnegative matrix factorization (NMF), prosthesis, synergy. I. INTRODUCTION P ROSTHETIC limbs generally available to amputees in- clude simple grasping mechanisms. A standard one-to-one mapping control paradigm typically includes two myoelectric sites to detect electromyogram (EMG) signals from flexor and extensor muscles on the forearm. Users can only control a single degrees of freedom (DOF) of the prosthesis with the activities of these muscles. Multiple-DOF control (switching between dif- ferent DOFs) needs to be realized through the incorporation of some additional commands (e.g., cocontraction). Modern pros- thetic hands have multiple fingers which offer higher DOF [1]. However, EMG-based control algorithms do not have compara- ble control capability [2]. Higher DOFs increase the functional complexity of the prosthesis and, hence, decoding subjects’ in- tent becomes difficult. Pattern recognition (machine learning) methods can enhance prosthesis control. In myoelectric prosthesis related studies, various pattern recognition algorithms have been used. Arti- ficial neural network [3], support vector machine [4], Bayesian Manuscript received January 7, 2014; revised July 26, 2014; accepted August 20, 2014. Date of publication December 11, 2014; date of current version January 13, 2015. This paper was recommended by Associate Editor R. A. Hess. J. Ma and F. Matsuno are with the Department of Mechanical Engineering and Science, School of Engineering, Kyoto University, Kyoto 6158530, Japan (e-mail: ma.jiaxin.62c@st.kyoto-u.ac.jp; matsuno@me.kyoto-u.ac.jp). N. V. Thakor is with the Department of Biomedical Engineering, School of Medicine, The Johns Hopkins University, Baltimore, MD 21205 USA (e-mail: thakorjhu@gmail.com). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/THMS.2014.2358634 network [5], the Gaussian mixture model classifier [6], and hidden Markov models [7] are some widely used algorithms. Decoding has also been performed for multiple DOF control cases, such as individual finger movement [8]. Through train- ing, a pattern recognition method can “learn” various states such as “open hand” or “close hand” and recognize them perfectly. However, a prosthesis does not have limited discrete outputs. Between “open hand” and “close hand,” there is a series of transition states. Proportional control of prostheses cares about how to recognize these transition states; however, training all of them is time consuming and unrealistic. Moreover, for a multiple-DOF prosthesis, even when each DOF is well trained, it does not mean that the user can simultaneously control two DOFs. For example, usually a classifier recognizes single states “open” and “pronate,” but cannot respond to the combination state “open and pronate” correctly. While regression methods are more appropriate for solving these problems due to their con- tinuous nature [9], [10], they still require supervised training. The training data must involve all of the transition states, and all of them must be correctly labeled. Such training processes are also complicated and time consuming. As a result, a method which can recognize untrained transition states or combination states, with less training time, is preferred. Neuroscience can support building models for natural move- ments, where a typical representation is the muscle synergy. Unlike normal pattern recognition methods which are based on statistical rules, this theory offers an understanding of the underlying processes of limb movements; therefore, it can po- tentially recognize untrained states. According to the theory of muscle synergies, when a subject’s central nervous system (CNS) controls the body movements, it does not send a single command to multiple muscles individually. Instead, it activates a small set of coordinated movements or a cascade of movement primitives called synergies. Each synergy is responsible for ac- tivating a larger set of muscles. The theory has been tested on frogs, mammals, primates, and humans [11]–[14]. An example of a marionette can also illustrate the basic idea: only several strings can control the complex movements of the puppet be- cause the manipulator’s hand (equivalent to the CNS) sends out higher order commands to activate the strings (synergies) and achieves control of multiple joints of the puppet (human limbs). Synergies and muscles possess many-to-many mappings, and the number of synergies needed is typically far less than the number of muscles. This approach of modeling limb motion can also be regarded as a dimension reduction mechanism from muscles to synergies. If the synergies can be calculated for an amputee, then it would be possible to represent the extent of a single movement, as well as the simultaneous activation of multiple movements. 2168-2291 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.