Emg Amplitude Estimation: a Review of the past and a Look towards the Future

AND INTRODUCTION The amplitude of the surfa.ce EMG is frequently used as the control input to myoelectric prostheses, as a measure of muscular effort, and has also been investigated as an indicator of muscle force This paper will review the methods which are used to estimate the EMG amplitude from recordings of the EMG waveform. (Note that this review does not include the related area of EMG-to-force processing.) Early investigators studied the type of non-linear detector which should be applied to the waveform This work led to the routine use of analog rectify and smooth (low pass filter) processing and root-mean-square (RMS) processing of the EMG wavefomi to form an amplitude estimate, More recent investigation has shown the prornise of whitening individual EMG waveform channels, combining multiple waveform channels into a single EMG amplitude estimate and adaptively tuning the smoothing window length None of these recent techniques have been routinely incorporated into EMG amplitude estimators Finally, a look towards what EMG processing might be in the future is described AMPLITUDE DETECTION Early investigators treated the EMG wavefbrm as a zero mean amplitude modulated signal Inman et al [11] suggested an amplitude estimator consisting of a full-wave rectifier (non-linear demodulator) followed by a simple resistor-capacitor low pass filter (smoother) They noted that a long filter time constant was desired to reduce noise in the estimate, while a short filter time constant reduces the time delay in tracking changes in the signal amplitude This simple detector, with time constants from 0 I to 1 0 second, has been applied extensively in the study of the surface electromyogram. In an effort to improve the estimator of Inman et al [11], Kreifeldt [13] compared the performance of three smoothers the standard resistor-capacitor. (RC) low pass filter, a thirdorder Butterworth filter, and a third-order approximation to a moving average filter. The performance of each amplitude estimator was evaluated from EMG data recorded during isometric, isotonic muscular contraction by computing a signal to noise ratio (SNR) from the output of each amplitude estimator . Because muscle contraction was constant, the mean value of an amplitude estimate was taken as the signal, and standard deviations about the mean were taken as noise (T'his performance criterion has been widely adopted in surface EMG amplitude estimation ) Kreifeldt found that the SNR performance of the averaging filter was a 44% improvement over the RC filter, while the Butterworth filter provided an 11% improvement. Kreifeldt and Yao [14] experimentally investigated the performance of six non-linear demodulators A second-power demodulator was found to be best for contraction levels of 10, 25 and 50% maximum voluntary contraction (MVC). A fourth-power demodulator was found to be best at 5% MVC. These power law demodulators improved the SNR performance of the full wave rectifier. by 5-20%, depending on the force level Hogan and Mann [9,10] used a functional mathematical model of EMG to analytically predict that a second-power demodulator and an averaging filter, i e an RMS processor, would give the best maximum likelihood estimate of the EMG amplitude, Experimentally, they confirmed that an RMS processor is superior in SNR performance to a low pass filter. by 26%. Hogan and Mann found no SNR performance difference between the RMS processor and a full wave rectifier.. Clancy [1] consistently found full wave rectification to be a small improvement (2-8%) over RMS detection. Typical EMG amplitude estimators in use today utilize one of the above specified processors, with RMS processing being preferred. WHITENING FILTERS Several investigators have found that the inclusion of a whitening filter prior to demodulation and smoothing improves the performance of the amplitude estimate. A whitening inotitute of Oiomezlicat Engineering, Univereity of New 5runewick, MEG '97 nissueB in Upper Limb Prosthetico" filter is a filter whose output power spectrum is constant-valued when presented with the signal of interest as an input Kaiser and Peterson [12] found that the shape of the whitening filter should change as a function of the contraction level. They suggested that measurement noise, present in differing relative degrees depending on the absolute signg (contraction) level, may be a major factor in determining the shape of the whitening filter,. They designed an adaptive analog filter to achieve their desired whitening. Harba and Lynn [8] used auto-regressive modeling of the EMG power spectrum to form a whitening filter in an off-line algorithm. Their sixth-order model found only small changes in the shape of the whitening filter as a function of the contraction level. Whitening approximately doubled the probability of correctly differentiating between one of four discrete contraction leVels Their off-line results were confirmed with an analog on-line implementation Hogan and Mann [9,10] found that whitening could be achieved by reducing the outer edge spacing of a pair of rectangular electrodes from 20rnm to lOmm An SNR performance improvement of 35% resulted D'Alessio [4] and Filligoi and Mandarini [7] discussed whitening with respect to functional mathematical models of the EMG,. Clancy and Hogan [2] systematically investigated the influence of various moving-average digital whitening filters for contractions over the range of 10-75% MVC, They found that fourth-order whitening filters, calibrated froma short segment (55s) of data, improved the SNR by 63% These whitening filters, however, performed poorly for contractions less than 10% MVC, Additive background noise seemed to dominate the output of the whitening filters, As in the prior work of Kaiser and Peterson [12], Clancy [1] implemented an adaptive whitening filter which seemed to maintain the SNR performance improvement for contractions above 10% MVC, but reverted towards unwhitened processing for lower levels of contraction MULTIPLE SITE COMBINATION Further improvements in EMG amplitude estimation have been achieved through the combination of multiple channels of the EMG waveform Hogan and Mann [9,10] suggested that dispersing multiple electrodes about a single muscle would provide a broader, more complete, measure of the underlying electrophysiologic activity. They derived an optimal amplitude estimator assuming tha separate EMG channels were spatially correlated but temporally uncorrelated Using four electrodes, they achieved an SNR performance improvement of 91% compared to the single channel estimator of Inman et al. [11] The combination of multiple channels and whitening via electrode geometry yielded an SNR performance improvement of 176% compared to the single channel estimator of Inman et al [11], The SNR performance of their algorithm was relatively insensitive to force levels over the range of 5-25% MVC Hogan and Mann implemented their algorithm off-line on a digital computer and on-line with analog circuitry. Murray and Rolph [16] implemented this algorithm in real time on a digital microprocessor. Harba and Lynn [8] used four electrode pairs to improve the quality of an EMG processor. which tried to differentiate between four discrete contraction levels. They were able to improve the probability of correctly differentiating between contraction levels by 40-70% (compared to using one electrode) Thesneyapan and Zahalak [18] reported a nine channel EMG amplitude estimator, Clancy and Hogan [3] combined the techniques of waveform whitening and multiple channel combination. For contractions ranging from 10-75% MVC, a four channel, temporally whitened processor improved the SNR 187% compared to the estimator of Inman et al [11] Eight whitened combined channels provided an SNR improvement of 309% compared to the estimator of Inman et al. Calibration of the optimal processor was achieved with a single five second contraction trial at 50% MVC NON-STATIONARY PROCESSORS In all of the EMG processors described above, selection of the smoothing window length (or time constant, as appropriate) is a trade-off between amplitude estimator variance (which is diminished via a long smoothing window) and error due to estimator bias (which is diminished via a short smoothing window) When EMG amplitude is varied dynamically during muscular contraction (i e. the EMG waveform is non-stationary), higher fidelity EMG amplitude estimates can be achieved if the window length is tuned throughout the duration of the contraction