REAL-TIME STIMULATION ARTIFACT REMOVAL IN EMG SIGNALS FOR NEUROPROSTHESIS CONTROL APPLICATIONS

An elegant and intuitive method to control a neuroprosthesis is to use S(urface)EMG activity of voluntary controllable muscles [1]. Since in such applications the voluntary SEMG activity is contaminated with much higher stimulation artifacts (SA) than the SEMG signal, the artifact somehow has to be eliminated. Well-established SA removal techniques are artifact blanking or filtering methods. Real-time SA blanking methods, either hardware sample-hold circuits or digital blanking routines loose all EMG information during the blanking period that lasts several milliseconds. SA filtering techniques are not practical with constant current stimulators since the long lasting SA tail overlaps in frequency and time domain with the voluntary SEMG. A new method that makes use of the randomness and stationarity of voluntary EMG is presented. An ensemble averaged SA with exponential forgetting was subtracted from the recorded SEMG and an almost artifact free SEMG signal was obtained. Fast convergence of the algorithm and good quality residual SEMG were shown, while the real-time computational power requirements were very low. Introduction A SEMG signal that is recorded during surface FES from a muscle close to the stimulation site is always contaminated with SA. Even if the stimulation site like in this study is far away (we stimulate the distal arm and record on the contralateral deltoid muscle), the SA is at least the double amplitude than the strongest recorded SEMG from voluntary muscle contraction (Figure 1, 3 curve). Close to the stimulation site the SA consists of (1) the stimulation spikes that drive the EMG amplifier into saturation; (2) a fast decaying artifact tail produced by the EMG AC-coupling filter; (3) and a slow decaying artifact tail produced by the slowly discharging electrode-tissue impedance (Figure 1, 1 curve). In case of a constant current stimulator that has very high output impedance, the decay of the long artifact tail can last longer than 10 ms. Many different methods that remove the SA from EMG or similar neurophysiological signals were proposed in the last 30 years. They can be divided in three main categories: SA blanking, SA filtering, and SA subtraction methods. Hardware [2, 3] and software [1, 4] artifact blanking or sample-and-hold blanking are simple techniques, that can be easily implemented in actual microcontrolled, electrical stimulators for real-time processing of SEMG signals. They blank or sample-and-hold the SEMG during the SA while loosing all signal information during that time. For low stimulation frequencies and few stimulation channels this technique can be applied to control a neuroprosthesis. But for higher stimulation frequencies or many stimulation channels the blanking time, especially with constant current stimulators, becomes too long and the SEMG signal looses its stationarity features. SA filtering methods [5-8] reduce the SA using linear, non-linear, or/and adaptive filtering, gain switching, slew rate limiting, or constant current/voltage switching techniques. They try to preserve more of the SA contaminated SEMG signal by reducing the SA spike (lowpass filters, slew rate limiters), by reducing the SA tail (gain or current/voltage switching methods) or by estimating the SA and filtering it (adaptive filter methods). But, because the SEMG signal and the SA overlap in time and frequency domain all applied filters influence the quality of the SEMG signal. The switching methods potentially cause transients and adaptive filters may have a slow convergence in the case the SA is changing like in FES applications. Software artifact subtraction methods [9-11] subtract a more or less pure SA from the mixed signal. The presented methods differ in the way the pure SA is obtained. Sub-motor-threshold stimulation, off-nerve recording, double-pulse stimulation within the refractory period of the nerve fiber, or ensemble averaging of the SA contaminated mixed signal are some of the presented methods. For the control of neuroprostheses the proposed SA subtraction algorithms cannot be used, because the produced SA changes with the action (e.g. grasping or releasing) over time. An a priori extracted SA cannot be adjusted to the measured SA in real-time during stimulation, since the changes of the SA are nonlinear (e.g. skin-tissue impedance changes) and depend on many unknown factors. To overcome the above problems an enhanced ensemble averaged SA subtraction method with real-time capabilities was developed. It adapts the subtracted SA to changes of the stimuli that occur in real FES applications. The performance of the SA removal algorithm was tested with stimulation patterns similar to the ones used in our grasping neuroprostheses [1]. Methods The SA was extracted from the first 12.5 ms post stimuli of the recorded SEMG signal. A moving ensemble averaging algorithm with exponential forgetting was used to extract the SA and the direct muscle responses, if present. The algorithm was on purpose kept very simple using a first order infinite impulse response (IIR) filter for the exponential forgetting. For each sample n the following recursive filter output was calculated: