Application of wavelet packet transform on myoelectric pattern recognition for upper limb rehabilitation after stroke.

Myoelectric pattern recognition applied to high-density surface electromyographic (sEMG) recordings from paretic muscles has been proven to identify various movement intents of stroke survivors, thus facilitating the design of myoelectrically controlled robotic systems for recovery of upper-limb dexterity. Aiming at effectively decoding neural control information under the condition of neurological injury following stroke, this paper further investigates the application of wavelet packet transform (WPT) on myoelectric feature extraction to identify 20 functional movements performed by the paretic upper limb of 4 chronic stroke subjects. The WPT was used to decompose the original sEMG signals via a tree of subspaces, where optimal ones were selected in term of the classification efficacy. The energies in the selected subspaces were calculated as optimal wavelet packet features, which were finally fed into a linear discriminant classifier. The WPT-based myoelectric feature extraction approach achieved accuracies above 94% for all subjects in a user-specific condition, demonstrating its potential applications in upper limb rehabilitation after stroke.