Classification of motor unit activity following targeted muscle reinnervation

For the past six decades, signal processing methods for myoelectric control of prostheses consisted mainly of calculating time- and frequency domain features of the EMG signal. This type of feature extraction considers the surface EMG as colored noise, neglecting its generation as a sum of motor unit activities. In this study we propose the use of motor unit behavior for classifying motor tasks with the aim of myoelectric control. We recorded high-density surface EMG of three patients who underwent targeted muscle reinnervation, and decomposed these signals into motor unit spike trains using an automatic offline EMG decomposition method. From the motor unit spike trains we used the number of discharges in each analysis interval as a feature for a support vector machine classifier. The same classifier was used for discriminating classic time-domain EMG features, for comparison. Classification accuracy was greater for motor unit information than for the classic features (97.06%±1.74 vs 85.01%±13.66), especially when the number of classes was high (95.11% ± 1.74 vs 69.25% ± 4.04 for 11 classes). These results suggest that the identification of motor unit activity from surface EMG can be a powerful way for pattern recognition in targeted muscle reinnervation patients.