IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 52, NO. 11, NOVEMBER 2005 1801 A Gaussian Mixture Model Based Classification Scheme for Myoelectric Control of Powered Upper Limb Prostheses Yonghong Huang, Kevin B. Englehart*, Senior Member, IEEE, Bernard Hudgins, Senior Member, IEEE, and Adrian D. C. Chan, Member, IEEE Abstract—This paper introduces and evaluates the use of Gaussian mixture models (GMMs) for multiple limb motion classification using continuous myoelectric signals. The focus of this work is to optimize the configuration of this classification scheme. To that end, a complete experimental evaluation of this system is conducted on a 12 subject database. The experiments examine the GMMs algorithmic issues including the model order selection and variance limiting, the segmentation of the data, and various feature sets including time-domain features and autoregressive features. The benefits of postprocessing the results using a majority vote rule are demonstrated. The performance of the GMM is compared to three commonly used classifiers: a linear discriminant analysis, a linear perceptron network, and a multilayer perceptron neural network. The GMM-based limb mo- tion classification system demonstrates exceptional classification accuracy and results in a robust method of motion classification with low computational load. Index Terms—Classification, EMG, Gaussian mixture model, myoelectric signals, pattern recognition, prosthesis. I. INTRODUCTION E LECTRICALLY powered prostheses with myoelectric control have been found to have many advantages over other types of prostheses, mostly due to the autonomous na- ture of control. The myoelectric signal (MES), which can be collected at the skin surface using electrodes, is capable of providing information about neuromuscular activity in a nonin- vasive manner and, thus, has become an important and effective control input for powered prostheses. Pattern recognition of the MES, to discriminate amongst the desired classes of limb activations, plays a key role in advanced control of powered prostheses for individuals with amputations or congenitally deficient upper limbs. The success of a myoelectric control scheme depends greatly on the classification accuracy. Feature extraction and classification methodologies are two crucial Manuscript received June 30, 2005; revised February 2, 2005. This work was supported in part by the Natural Sciences and Engineering Research Council of Canada under Discovery Grant 217354-01 and Discovery Grant 171368-03. Asterisk indicates corresponding author. Y. Huang and B. Hudgins are with the Department of Electrical and Computer Engineering and the Institute of Biomedical Engineering, University of New Brunswick, Fredericton, NB E3B5A3, Canada. *K. B. Englehart is with the Department of Electrical and Computer Engineering and the Institute of Biomedical Engineering, University of New Brunswick, Fredericton, NB E3B5A3, Canada (e-mail: kengleha@unb.ca). A. D. C. Chan is with the Department of Systems and Computer Engineering, Carleton University, Ottawa ON K1S5B6, Canada. Digital Object Identifier 10.1109/TBME.2005.856295 factors for achieving high classification performance in pattern recognition. The concept of employing pattern recognition for MES con- trol is by no means new; indeed, the first pattern recognition based control schemes were developed as early as the late 1960s and early 1970s [1]–[3]. These efforts were encumbered by the limited computing capacity of the day, and the rather bulky instrumentation for multichannel acquisition; real-time imple- mentation was simply not feasible. Significant advances in pat- tern recognition methodology were yet to come, including artifi- cial neural networks and modern statistical classifiers, including hidden Markov models and Gaussian mixture models (GMMs). The first pattern recognition based approach that offered real- time performance and high accuracy was that by Hudgins et al. [4]. This approach was based upon a set of simple time domain (TD) statistics and a multilayer perceptron (MLP) neural net- work classifier; the system was capable of classifying four types of limb motion with an error rate of around 10% [4]. This work incited a renewed interest in pattern recognition based MES con- trol, and a great deal of work ensued. These efforts have in- vestigated the efficacy of various feature sets [5]–[8] (mostly time-frequency and time-scale representations), and classifiers, including dynamic artificial neural networks [9], genetic algo- rithms [8], fuzzy logic classifiers [10], [11], and self-organizing neural networks [12]. These systems show great promise, as they exhibit high ac- curacy, and are capable of being tuned to an individual user, learning the characteristics of the MES activity accompanying their contraction efforts. The current authors have recently de- scribed a practical system which is exceptionally accurate, and can classify a continuous stream of myoelectric activity in real- time [13], [14]. Moreover, this approach can exploit the infor- mation in an arbitrary number of MES channels, and is immune to crosstalk that may exist between these recording sites. Indeed, the crosstalk constitutes information regarding the spatial-tem- poral relationship in the MES activity amongst the channels, and may be useful in the classification process. Chan et al. extended this work, demonstrating the potential of the GMM in MES classification [15]. This research builds upon Chan’s preliminary work and optimizes the GMM approach for multiple limb motion classification using the MES. The GMM has become the dominant approach in speaker recognition and verification over the past several years [16]–[19]. The GMM has the ability to form smooth ap- proximations for general probability density functions through 0018-9294/$20.00 © 2005 IEEE Authorized licensed use limited to: Oregon Health & Science University OHSU. Downloaded on May 27,2010 at 17:21:05 UTC from IEEE Xplore. Restrictions apply.