Appropriate AR modeling for surface electromyogram signals and its application in hand activity classification

Abstract

Surface electromyogram (sEMG) has found wide range of applications in human machine interface, assistive technology and health monitoring. A simple autoregressive (AR) model may be used to describe the shape of the signal spectrum. Then, the main concern is the manner in which the residual signal of the AR model is parameterized. It has been shown only recently that the sEMG signals exhibit heteroscedasticity resulting in the AR residual signal being heteroscedastic. In this paper, the aim is to explore the effect of using different order AR models with different residual signal models on sEMG-based classification of hand activities. It is demonstrated that the appropriateness of the AR model order should be determined by jointly testing the AR and residual model parameters for classification in terms of the accuracy that they provide. A stand-alone statistical test for determining the AR model order may not correspond with the accuracy that the model would provide when used in conjunction with the features extracted from the residual signals. Moreover, feature selection is essential while testing large number of features to determine the appropriateness of the signal models used.