Electromyogram (EMG) Signal Analysis: Extraction of a Novel EMG Feature and Optimal Root Difference of Squares (RDS) Processing in Additive Noise

Kiriaki J. Rajotte, Worcester Polytechnic Institute

Abstract

Electromyogram signals generated by human muscles can be measured on the surface of the skin and then processed for use in applications such as prostheses control, kinesiology and diagnostic medicine. Most EMG applications extract an estimate of the EMG amplitude, defined as the time-varying standard deviation of EMG, EMGσ. To improve the quality of EMGσ, additional signal processing techniques, such as whitening, noise reduction and additional signal features can be incorporated into the EMGσ processing. Implementation of these additional processing techniques improve the quality of the processed signal but at the cost of increased computational complexity and required calibration contractions. Whitening filters are employed to temporally decorrelate data so that the samples are statistically independent. Different types of whitening filters, linear and adaptive, and their performance have been previously studied in (Clancy and Hogan) and (Clancy and Farry). The linear filter fails at low effort levels and the adaptive filter requires a calibration every time electrodes are removed and reapplied. With the goal of avoiding the disadvantages of the previous whitening filter approaches, the first signal processing technique studied herein developed a universal fixed whitening filter using the ensemble mean of the power spectrum density of EMG recordings from the 64 subjects available in an existing data set. Performance of the EMG to torque model with the universal fixed whitening filter was computed to be 4.8% maximum voluntary contraction (MVC); this is comparable to the 4.84 %MVC error computed for the adaptive whitening filter. The universal fixed whitening filter preserves the performance of the adaptive filter but need not be calibrated for each electrode. To optimize noise reduction, the second signal processing technique studied derived analytical models using the resting EMG data. The probability density function of the rest contractions was observed to be very close to a Gaussian distribution, showing only a 1.6% difference when compared to a Gaussian distribution. Once the models were developed, they were used to prove that the optimal subtraction of the noise variance is to compute the root of the difference between the signal squared and noise variance (RDS). If this result would lead to a negative value, it must be set to zero; EMGσ cannot contain negative components. Once the RDS was proven to be the optimal noise subtraction, it was implemented on 0% MVC and 50% MVC data. The RDS processing has a considerable impact on lower level contractions (0% MVC), but not on higher level contractions (50% MVC), as expected. The third signal processing technique involved the creation of a new EMG feature from four individual signal features. Different techniques were used to combine EMGσ, zero crossings (ZC), slope sign changes (SSC) and waveform length (WL) into a single new EMG feature that would be used in an end application, such as the modeling of torque about the elbow or prosthesis control. The new EMG feature was developed to reduce the variance of the traditional EMGσ only feature and to eliminate the need for calibration contractions. Five different methods of combination were attempted, but none of the new EMG features improved performance in EMG to torque model.