Pattern classification of hand movements using time domain features of electromyography

Myoelectric control of prostheses is a long-established technique, using surface electromyography (sEMG) to detect the electrical signals of muscle activity and perform subsequent mechanical actions. Despite several decades’ research, robust, responsive and intuitive control schemes remain elusive. Current commercial hardware advances offer a variety of movements but the control systems are unnatural, using sequential switching methods triggered by specific sEMG signals. However, recent research with pattern recognition and simultaneous and proportional control shows good promise for natural myoelectric control. This paper investigates several sEMG time domain features using a series of hand movements performed by 11 subjects, taken from a benchmark database, to determine if optimal classification accuracy is dependent on feature set size. The features were extracted from the data using a sliding window process and applied to five machine learning classifiers, of which Random Forest consistently performed best. Results suggest a few simple features such as Root Mean Square and Waveform Length achieve comparable performance to using the entire feature set, when identifying the hand movements, although further work is required for feature optimisation.