Assessing fractal dimension methods as feature extractors for EMG signal classification

The study of electromyographic (EMG) signals has gained increased attention in the last decades since the proper analysis and processing of these signals can be instrumental for the diagnosis of neuromuscular diseases and the adaptive control of prosthetic devices. As a consequence, various pattern recognition approaches, consisting of different modules for feature extraction and classification of EMG signals, have been proposed. In this paper, we conduct a systematic empirical study on the use of Fractal Dimension (FD) estimation methods as feature extractors from EMG signals. The usage of FD as feature extraction mechanism is justified by the fact that EMG signals usually show traces of self-similarity and by the ability of FD to characterize and measure the complexity inherent to different types of muscle contraction. In total, eight different methods for calculating the FD of an EMG waveform are considered here, and their performance as feature extractors is comparatively assessed taking into account nine well-known classifiers of different types and complexities. Results of experiments conducted on a dataset involving seven distinct types of limb motions are reported whereby we could observe that the normalized version of the Katz's estimation method and the Hurst exponent significantly outperform the others according to a class separability measure and five well-known accuracy measures calculated over the induced classifiers.