Surface myoelectric signal analysis: Dynamic approaches for change detection and classification

Toward the goal of elbow and wrist prostheses control by characterizing events in surface myoelectric signals, this paper presents a dynamic method to simultaneously detect and classify such events. Dynamic cumulative sum of local generalized likelihood ratios using wavelet decomposition of the myoelectric signal is used for on-line detection. Frequency as well as energy changes are detected with this hybrid approach. Classification is composed of using multiresolution wavelet analysis and autoregressive modeling to extract signal features while polynomial classifiers are used for pattern modeling and matching. The results of detecting and classifying four elbow and wrist movements show that, in average, 91% of the events are correctly detected and classified using features obtained from multiresolution wavelet analysis while 95% accuracy is achieved with AR modeling. The classification accuracy decreases, however, if short prostheses response delay is desired. This paper also shows that the performance of the polynomial classifiers is better than that of the commonly used neural networks since it gives higher classification accuracy and consistent classification outcomes. In comparison to the well known support vector machine classification, the polynomial classifier gives similar results without the need to optimize and search for classifier parameters.