Dynamic Behavior of Time-Domain Features for Prosthesis Control

Myoelectric hand-prostheses are used by patients with either above- or below-elbow amputations and actuated with a minimal microvolt-threshold myoelectric signal (MES). Prehensile motions or patterns are deduced from the MES by classification. Current approaches act on the assumption, that MES is adiabatic-invariant and unaffected by fatigue of contributory muscles. However, classifiers fail on the onset of muscle fatigue and cannot distinguish between voluntary-, submaximal contraction and in intentional release of muscle tension. As a result, patients experience, a, gradual loss of control over their prostheses. In this coutribution we show, that the probability distributions of extracted time- and frequency-domain features are fatigue dependent with regard to locality, skewness and time. Also, we examine over which time-frame, established classifiers provide unambiguous results and how classifiers can be improved by the selection of a proper sampling-window size and an appropriate threshold for select features.