1762 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 10, OCTOBER 2007 Application of Higher Order Statistics to Surface Electromyogram Signal Classification Kianoush Nazarpour, Student Member, IEEE, Ahmad R. Sharafat*, Senior Member, IEEE, and S. Mohammad P. Firoozabadi Abstract—We propose a novel approach for surface electromyo- gram (sEMG) signal classification. This approach utilizes higher order statistics of sEMG signal to classify four primitive motions, i.e., elbow flexion, elbow extension, forearm supination, and forearm pronation. In documented research, the sEMG signal generated during isometric contraction is modeled by a stationary process whose probability density function (pdf) is assumed to be either Gaussian or Laplacian. In this paper, using Negentropy, we demonstrate that the level of non-Gaussianity of sEMG signal recorded in muscular forces below 25% of maximum voluntary contraction (MVC) is significant. Therefore, application of higher order statistics in sEMG signal processing is justified, due to the fact that more useful information can be extracted from the corresponding higher order statistics. An accurate classifica- tion is achieved by using the sequential forward selection (SFS) method for reducing of the dimensionality of feature space and the -nearest neighbor (KNN) classifier. The results indicate that the proposed approach provides higher sEMG correct classification rates as compared to the existing methods. Index Terms—Higher order statistics, negentropy, sequential forward selection, surface electromyogram signal. I. INTRODUCTION S URFACE electromyogram signal is the electric manifesta- tion of neuromuscular activities [1] and is recorded nonin- vasively from the skin by using biopotential electrodes [2]. It is an intricate signal that depends on the anatomical and physio- logical properties of the contracting muscles beneath the skin [1]. The sEMG signal has been widely applied in rehabilitation and control of prosthetic devices for individuals with amputa- tions or congenitally deficient limbs [3]. The control mecha- nism is based on correct classification of sEMG signals recorded during muscular contraction. Various techniques have been employed for sEMG signal processing, such as autoregressive (AR) modeling [3]–[5], statistical pattern recognition [6], discrete wavelet transform Manuscript received January 2, 2006; revised January 5, 2007. This work was supported in part by Tarbiat Modares University, Tehran, Iran. Asterisk indicates corresponding author. K. Nazarpour was with the Department of Electrical and Computer Engi- neering, Tarbiat Modares University, Tehran, Iran. He is now with the Centre of Digital Signal Processing, School of Engineering, Cardiff University, Cardiff CF24 3AA, U.K. (e-mail: NazarpourK@cf.ac.uk). *A. R. Sharafat is with the Department of Electrical and Computer En- gineering, Tarbiat Modares University, P. O. Box 14155-4838, Tehran, Iran (e-mail: Sharafat@isc.iranet.net). S. M. P. Firoozabadi is with the Department of Medical Physics, Tarbiat Modares Univerity, Tehran, Iran. (e-mail: pourmir@modares.ac.ir). 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/TBME.2007.894829 (DWT) and wavelet packet transform (WPT) [7], and artificial neural network architectures together with other feature extrac- tion schemes [8]–[10]. In many sEMG classification schemes, the time-domain features such as the mean absolute value, the Wilson amplitude, the zero crossing, and the autoregressive co- efficients of the sEMG signal, as well as the frequency-domain features such as the amplitude of the spectrum of the windowed sEMG in selected bands are applied to the classifier. In [6], a considerable set of these features have been evaluated in different time window sizes and compared for the classifica- tion rate, robustness to noise, and computational complexity. Although the existing methods attain high rates of correct classification, substantial computations are entailed. The sEMG classification involves several stages, namely sEMG detection, motion class formation, feature extraction, sEMG classifica- tion, and error estimation. As correct classification depends on extracting distinctive features [11], we focus on extracting such features from higher order statistics of the sEMG recordings from biceps brachii and triceps brachii muscles to identify four primitive motions, i.e., elbow flexion, elbow extension, forearm supination, and forearm pronation. The impetus behind this study is to establish the applicability of a selective combination of higher order statistics in the sEMG signal classification. It has long been held that the sEMG signal recorded during constant-force, constant-angle, and nonfatiguing contractions can be modeled as a zero-mean Gaussian stationary process [12]. In [13], it is affirmed that the Gaussian distribution accu- rately describes the density for various contraction strengths, and for the biceps sEMG signal, using Chi-square test, the probability of deviation from Gaussian distribution was less than 10 . Although the pdf of the sEMG is presumed to be Gaussian, in some cases it has been reported that the above pdf is closer to Laplacian, depending on the level of MVC (or the number of active motor units). In [14], it is reported that during isometric contractions, the pdf of the sEMG signal is more sharply peaked near zero than the Gaussian distribution. In [15], the pdf of the sEMG signal recorded from biceps in constant forces (20%, 40%, 60%, and 80%) was reported to be non-Gaussian, which moves towards Gaussian in higher force levels. In some recent studies [16]–[21], the non-Gaussian signal processing schemes for sEMG analysis have been devel- oped. In [18], a higher order statistics (HOS)-based technique is utilized to characterize the motor unit action potentials (MUAPs), where conventional approaches are generally based on the analysis of the first- and the second-order moments and cumulants (i.e., mean, correlation, and variance) and their spectral representations. Such techniques assume that the 0018-9294/$25.00 © 2007 IEEE