A Comparative Study of the Techniques for Decomposition of EMG Signals

Abstract

This paper deals with four decomposition algorithms, which have been modified, implemented, analyzed and evaluated, for their performance in separation of motor unit action potentials (MUAPs) from the Electromyogram (EMG) Signals. The performance of algorithms has been evaluated to determine, as to which one out of the four algorithms is accurate, fast, reliable, efficient and can extract clean MUAPs even from those EMG signals which have been recorded for limited duration. Both synthetic and real time EMG signals have been used for testing the algorithms. The classification success rate achieved with statistical pattern recognition and cross-correlation approaches is 98.9% and 98.8% respectively whereas with Kohonen neural network 99.2% and wavelet transform 99.8%. Therefore the wavelet transform method is recommended because of its highest success rate, as this method does not require any correction for baseline drift or high frequency moise. It allows fast extraction of the localized frequency components, provides good time-resolution, and is capable of tracking rapid changes in MUAPs. The superimposed signal, which could not be separated by one of the above technique, has been decomposed by using cross-correlation and Euclidean distance. For earlier three techniques, the results are given only in tabular form while for wavelet technique, the results are presented both in tabular and graphical forms. All the algorithms have been successfully implemented and tested for decomposition of EMG signals recorded from subjects having normal (NOR) state of muscles and having motor neuron disorder (MND) and myopathy (MYO) disease.