Optimal myoelectric feature space for pattern discrimination

The objective of this work was to find the optimal representation of the myoelectric (EMG) signal for discrimination of distinct muscle contraction patterns. One able-bodied subject, one traumatic below-elbow amputee, and four subjects with below-elbow congenital limb deficiencies constituted the subject population. The raw EMG signal was measured from two clinically chosen electrode sites on the subjects' residual limbs while a variety of contraction patterns (four or six depending on subject ability) were performed. Contractions were performed under simulated conditions of normal use. The EMG transient was extracted and a variety of feature space representations were calculated. An artificial neural network was trained and tested using each feature space in order to determine which performed best as a classifier of the distinct muscle contraction patterns. The mean rectified EMG, rectified EMG variance, and EMG power spectrum estimates were found to have similar classification properties. Combinations of features were found to yield no improvement in classification accuracy. Best classification results were obtained using either the mean rectified EMG or the rectified EMG variance calculated for three 70 ms time windows over the EMG transient. The use of six contraction patterns yielded poor classification results under all circumstances. Classification of four contraction patterns (five state system) yielded accuracy rates of approximately 70% using the rectified mean EMG. The author has contributed data which provides (a) a guide for selecting the optimal myoelectric feature space for developing a multiple function prosthesis when using clinically determined two-site myoelectric control. (b) the number of distinct contraction patterns that clients with a variety of limb deficiency lengths and types are capable of generating.