Effects of electromyographic processing methods on computer-averaged surface electromyographic profiles for the gluteus medius muscle.

Effects of Electromyographic Processing Methods on Computer-Averaged Surface Electromyographic Profiles for the Gluteus Medius Muscle During the past decade various reports [1-4] have described the use of computer-averaged electromyographic (EMG) profiles in the analysis of pathological gait. For instance, Knutsson [2] used computer-averaged EMG profiles to study muscle coordination in hemiparetic gait, Conrad et al [3] reported on the use of EMG profiles for investigating the gait patterns of paraspastic patients with a variety of diagnoses, and Winter and Sienko [4] used EMG profiles in an attempt to elucidate the mechanics of gait in patients with below-knee amputation. These studies show that computer-averaged EMG profiles provide potentially valuable information about the time course of myoelectric activity of muscles.

A muscle's EMG profile may be observed during a number of gait cycles, with the beginning of each new gait cycle defined by the moment of heel-strike. This "raw" EMG signal is usually full-wave rectified and low-pass biltered, an operation known as linear envelope detection. [5] The output signal of the detector represents the time course of the estimated intensity of the EMG signal and is stored in a computer to allow the second processing step. The intensity of the EMG signal, sometimes called the "amount of activity," [5] indicates the muscle's myoelectric activity. The accuracy of the estimate in this first step is governed by the bandwidths of the raw EMG signal and the low-pass filter used [6] and is limited. In the second step, the time scale of the recorded intensity signal in each individual gait cycle is normalized to a standard cycle length. The resulting intensity signal frames, which correspond to the individual cycles, are ensemble-averaged to form the final EMG signal intensity profile. This ensemble-averaging process improves the accuracy of the intensity information, because random fluctuations in the contributing signal frames tend to cancel each other. The choice of the low-pass filter in the linear envelope detector is an important consideration for the application of the EMG profiles. A wide variety of low-pass filters have been reported in the literature. Knutsson [2] reported time constants of both 0.2 and 0.02 second. Winter [1] used a critically damped filter with a cutoff frequency of 3 Hz. Arsenault et al [7] worked with a 6-Hz filter. Kleissen et al [8] used a third-order Butterworth filter with a cutoff frequency of 25 Hz. This lack of uniformity indicates that choosing a low-pass filter is not a trivial matter. It also makes comparison of the published results of different investigators difficult.

The purpose of this study was to demonstrate how the choice of the low-pass filter used in the linear envelope detector can affect the recorded EMG profiles. The results of this study may contribute to clinicians' understanding of the effects of EMG processing methods on computer-averaged EMG profiles.

Method

Instrumentation

In this study, computer-averaged profiles were recorded for the surface EMG activity of the gluteus medius muscle. Figure 1 schematically shows the instrumentation used for obtaining the EMG profiles. A K-Lab SPA-10 skin-mounted bipolar preamplifier, (*1) which snap-connects to two Medi-Trace self-adhesive disposable silver-silver chloride pellet electrodes (*2) picked up and amplified the EMG signal at the electrode location. This preamplifier provides a gain factor of 100 ([+ or -] 1%) and has an input impedance greater than 100 M[Omega], a common mode rejection ratio greater than 100 dB, and a passband of -3 dB between 20 Hz and 10 kHz. The rim-to-rim electrode distance was 22 mm, and the pickup area of each electrode was 10 X 10 [mm.sup.2]. The low mass of the amplifier (15 g) and its low output impedance reduce the problems of movement and cable artifacts. A reference electrode was attached to the wrist of the subject.

The time of heel-strike was identified using footswitches. The footswitch system comprised four sheets of conductive rubber, covering the forefoot and the heel of the shoes, in conjunction with a conductive rubber mat. When one of these sheets touched the rubber mat, an electric signal indicated floor contact.

A cable transported the preamplified EMG and footswitch signals to the data-processing station where they were conditioned before computer lrocessing. An electronic circuit removed occasional instability in the footswitch signals ("switch-bouncing") at heel-strike and toe-off. Processed signals representing the stance and swing phases of the gait cycle changed state only after the raw signals from the footswitches had been stable for 100 msec. The computer system took this 100-msec time delay into account.

The EMG signal first passed through a third-order high-pass Butterworth filter with a cutoff frequency of 20 Hz ([+ or -] 5%) for suppression of remaining movements and cable artifacts. This cutoff frequency specifies the frequency at which the filter's attenuation has increased to 3 dB. Subsequently, after further amplication and full-wave rectification, the EMG signal passed through two different low-pass filters connected in parallel. Filter A was a third-order Butterworth filter with a cutoff frequency of 25 Hz ([+ or -] 5%). Filter B was a critically damped second-order filter with two time constants of 47 msec ([+ or -] 10%). This filter is identical to two first-order Butterworth sections connected in series, each section with a cutoff frequency of 3.4 Hz ([+ or -] 10%). This filter will be referred to as the 3-Hz filter. Footswitch signals and the EMG signal, thus processed with two linear envelope detectors having only different low-pass filters, were sampled at 200 Hz and converted automatically into computer-averaged EMG profiles using the procedure and equipment described by Kleissen et al. [8]

On the conductive rubber mat, two infrared beams demarcated a 10-m walkway. Crossing the beams automatically started and stopped data collection of footswitch and EMG signals. An HP 3310B signal generator (*1) and a small loudspeaker served as a metrononme. Off-line data analysis was performed with a Tulip AT Compact 2 computer (*2) using the LOTUS 1-2-3 spreadsheet package. (*3)

Protocol

Eight healthy male subjects, aged 24 to 32 years, participated in this study. Each subject provided informed consent lrior to participation in the study. After thoroughly rubbing the skin over the gluteus medius muscle with alcohol, and after shaving the area when necessary, the electrodes were placed vertically on the muscle bellyof the gluteus medius muscle, halfway between the iliac crest and thegreater trochanter of the right leg. Before measurements were taken, the subjects were given a few minutes to practice the walking procedure so that they could become familiar with the experimental constraints. They practiced the procedure until they said they were steady in their walking pattern.

Each subject performed three walking trials. The first trial was at a comfortable, bree-walking cadence. During the second and third trials, the metronome was set at 78 and 120 beats per minute (bpm), respectively, and the subjects were instructed to adjust their walking cadence accordingly.

The subjects were free to choose their own comfortable stride length. An assistant walking behind the subject carried the trailing signal cable to avoid obstruction by cable drag.

Data were collected during 20 to 25 gait cycles. Because the 10-m walkway did not allow this number of gait cycles to be observed in one pass, the subjects were instructed to walk back and forth across the walkway until the desired number was reached.

Data Analysis

The computer system automatically determiend both the average EMG signal intensity profile of each trial and its standard deviation. Each stride period between two consecutive heelstrikes was normalized to 100% of gait-cycle duration, where 0% corresponded to heel-strik. The ouput signal of the linear evelope detector was interpolated at 0.125% gait-cycle intervals; thus, the normalized gait cycle contained 800 data points. Averaing the EMG data across the normalized strides for each of the 800 data points produced the ensemble-averaged profile. Similarly, the standard deviation for each of the data points of the final EMG signal intensity profile was calculated.

In off-line analysis, across-subject ensemble averages and their standard deviations were computer by averaging the average EMG profiles of the eight subjects. For a given ensemble average and its standard deviation, the coefficient of variation (CV) describes the variability. The CV is defined as the root-mean-square standard deviation over the stride period divided by the mean ensemble average over the stride. [1] Pearson's product-moment correlation coefficient (r) can be used to quantify the similarity of two EMG profiles. [9] The 800 data points defining one EMG profile are paired with the corresponding 800 data points form the other EMG profile. A high correlation coefficient indicates that the time courses of both EMG profiles are similar.

Because this coefficient does not necessarily express absolute differences between the EMG profiles, an additional measures for similarity (S), defined as the CV for the ensemble averages of the two EMG profiles to be compared, was computed. Statistical significance was set at the alpha level of .05.

Results

Figure 2 shows a typical result of one walking trial at a free-walking cadence. In this figure, A and B present the ensemble averages and standard deviation of EMG signal intensity for the gluteus medius muscle, as observed using a 25-Hz and a 3-Hz filter, respectively. It is evident that the 3-Hz filter produces a smoother EMg profile than the 25-Hz filter. The variability, as indicated by the standard deviation bars, also was lower for the 3-Hz filter than for the 25-Hz filter. Table 1 summarizes this effect of cutoff frequency on the variability in the EMG profiles of the individuals trials. Averages and standard deviations for the CV were calculated over the group of eight subjects. Table 1 indicates that the CV for the 3-Hz-filtered EMG profiles was lower than the CV for the 25-Hz-filtered EMG profiles by a factor of two, which was statistically significant (P [is less than] .05, Student's paired t test).

In Table 2, the similarity of the EMG profiles recorded at 78 and 120 steps/min for each subject is indicated by the average value and standard deviation of correlation (r) and similarity (S) over the group of eight subjects. Comparison of the values for correlation (r) and similarity (S) of the 25-Hz-filtered EMG profiles with those of the 3-Hz-filtered EMG profiles for each subject revealed that values for correlation (r) were significantly higher and that the values for similarity (S) were significantly lower for the 3-Hz-filtered EMG profiles (P [is less than] .05, Student's paired t test).

The results of pooling the EMG profiles across all subjects for each walking cadence are presented in Figure 3. In this figure, A and B present these across-subject averaged EMG profiles for the cut-off frequencies of 25 and 3 Hz, respectively.

In an attempt to quantify the changes in across-subject averaged EMG profiles with changing stepping frequencies, the similarity of the EMG profiles for 78 and 120 steps/min was parametrized with their correlation (r) values. The similarity (S) values for the averaged pair of these EMG profiles were also calculated. These values are presented in Table 3.

The CVs for the across-subject averaged AMC profiles for all walking cadences and filter frequencies are presented in Table 4. For the 3-Hz filter, CVs for these EMG profiles tended to decrease with increasing stepping frequency, whereas for the 25-Hz filter, the CVs tended to increase. Retrospectively, the average walking speeds were 0.83 (SD = 0.06), 1.41 (SD = 0.09), and 1.52 (SD = 0.08) m/sec for 78 steps/min, free-walking cadence, and 120 steps/min, respectively. The average stepping frequency over the group of eight subject at a free-walking cadence was 109 steps/min (SD = 4).

Discussion

Figure 2 shows that for a given raw AMC signal, the type of low-pass filter used in the linear envelope detector considerably affects the resulting EMG profiles. It is not surprising that the 3-Hz filter produces a smoother EMG profile than the 25-Hz filter, because it damps rapid fluctuations in the

EMG signal intensity estimate more strongly. Additionally, the 3-Hz-filtered EMG profiles lagged behind the 25-Hz-filtered EMG profiles. At a free-walking cadence, the peak in the EMG profile shown in Figure 2B appears approximately 50 msec later than the peak shown in Figure 2A. Filter theory states that the cutoff frequency and the time lag between filter input and output signal are closely related. Approximations show that this time lag would be inversely proportional to the cutoff frequency. [10]

The two different filters give two different representations of the same raw EMG signal. For the interpretation the EMG profiles, it is important to be aware of what is represented. Kleissen et al [8] viewed the linear envelope detector as an estimator of the intensity of the muscle activation that precedes the actual muscle contraction and force generation. The EMG profiles should provide detailed information about the effects of the motor control system on the muscle studied. The requirement that rapid fluctuations in the EMG signal intensity were to be observed led to the choice of the 25-Hz filter.

Winter [1] proposed that the linear envelope detector be designed so that its output will be closely related to the muscle force resulting from the observed EMG activity. He recognized that the time constants of a second-order critically damped low-pass filter could be tuned to achieve a dynamic input-output relationship that would parallel the dynamics of the EMG activity-force relationship of the muscle studied. For isometric contractions, there is both a theoretical basis [11] and experimental evidence [12] [13] to support this approach.

The effect of the low-pass filter on the individual EMG profiles propagates to the across-subject averaged EMG profiles shown in Figure 3. With increasing stepping frequency, the peak in the 25-Hz-filtered EMG profile shown in Figure 3(A) tends to grow and the EMG activity tends to concentrate progressively nearer to the time of heel-strike. In the 3-Hz-filtered EMG profile shown in Figure 3(B), this tendency is less distinct. The sensitivity of across-subject averaged EMG profiles for changes in the stepping frequency appears to be higher when filtered at 25 Hz. Data presented in Table 3 support this finding quantitatively. These findings confirm reported results from similar experiments. [14]

The across-subject averaged 3-Hz-filtered EMG profile for the gluteus medius muscle at a free-walking cadence agrees with those published by other investigators. [15,16] A review of the literature indicates that no EMG profiles for this muscle at other stepping frequencies have yet been published. The variability in the EMG profiles observed by Winter[15] (CV = 105%) is considerably higher than the variability observed in this study. Possible explanations may be a different electrode location used or a less homogeneous group of subjects.

Finally, it seems remarkable that the variability in the across-subject averaged 3-Hz-filtered EMG profiles decreases with increasing stepping frequency (Tab. 4), whereas this variability tends to increase for the 25-Hz-filtered EMG profiles. This finding means that the 3-Hz-filtered EMG profiles for the gluteus medius muscle of healthy individuals tend to become more similar with increasing stepping frequency.

The low-pass filter in real time forms the output signal as a moving average[5] over the rectified EMG signal where the shape of the averaging window is described by the filter's impulse response. However, if the filter's input signal is a burst in which the length is comparable to or shorter than the length of this averaging window, its output signal will be a smoothed version of the impulse response. With decreasing length of the input burst, the output signal will increasingly approach the fulter's impulse response. Figure 3A indicates that, with increasing stepping frequency, the gluteus medius muscle's EMG bursts become shorter. Therefore, the filter's response to the bursts will increasingly be determined by its impulse response, and the response will be more independent of the form of the input bursts. The response of the filter to the EMG bursts across different subjects will become more similar, so that the similarity of the individual subjects' EMG profiles can also increase with the stepping frequency. Because the impulse response of the 25-Hz filter was short compared with the length of the raw EMG bursts, this effect does not appear in Table 4 for the 25-Hz-filtered EMG profiles. At 120 steps/min, the EMG peak in Figure 3(A) lasted approximately 200 msec; the length of the impulse responses of the 25- and 3-Hz filters was approximately 35 and 200 msec, respectively.

Conclusion

Generally, the choice of the low-pass filter used in the linear envelope detector considerably affects both the shape and the properties of computer-averaged EMG profiles. In this report, some effects on the EMG profiles for the gluteus medius muscle have been described as an example of this general principle. The results of this study show that EMG profiles recorded with a linear envelope detector with a low cutoff frequency generally are smoother and have a lower cycle-to-cycle variability and a greater time lag with respect to the raw EMG signal than a filter with a higher cutoff frequency. The interpretation of the EMG profiles also is closely related to the type of filter used. The observed intersubject variability in the EMG profiles can be influenced by the low-pass filter: The longer the impulse response of the filter and the shorter the EMG bursts to be observed, the more the filter will shape the resulting EMG profile and the lower the intersubject variability will appear.

This study provides evidence that careful analysis of the nature of the information to be extracted from the EMG profile should precede the choice of the most suitable linear envelope detector for the particular problem under investigation. Standardization of the linear envelope detector, however, would be a contribution to avoiding confusion and misunderstanding and would allow better communication among various users of the EMG profiles. Further research may answer the question of whether a standard linear envelope detector that is suitable for a wide range of problems can be developed.

(*1) K-Lab, PO Box 70167, 1007 KD Amsterdam, the Netherlands.

(*2) Graphics Controls Canada Ltd, Gananoque, Ontario, Canada K7G 2Y7.

(*3) Hewlett-Packard Co, Intercontinental Headquarters, 3495 Deer Creek Rd, Palo Alto, CA 94304.

(*4) Tulip Computers, PO Box 3333, 5203 DH Den Bosch, the Netherlands.

(*5) Lotus Development Co, 55 Cambridge Pkwy, Cambridge, MA 02142.

References

[1] Winter DA. Pathological gait diagnosis with computer-averaged electromyographic profiles. Arch Phys Med Rehabil. 1984;65:393-398.

[2] Knutsson E. Gait control in hemiparesis. Scand J Rehabil Med. 1981;13:101-108.

[3] Conrad B, Benecke R, Meinck HM. Gait disturbances in paraspastic patients. In: Delwaide J, Young RR, eds. Clinical Neurophysiology in Spasticity. Amsterdam, the Netherlands: Elsevier Science Publishers BV; 1985;155-174.

[4] Winter DA, Sienko SE. Biomechanics of below-knee amputee gait. J Biomech. 1988;21:361-367.

[5] Winter DA, Rau G, Kadefors R, et al. Units, Terms and Standards in the Reporting of EMG Research. Montreal, Quebec, Canada; International Society of Electrophysiological Kinesiology (report of ad hoc committee); August 1980.

[6] Kadefors R. Myo-electric signal processing as an estimation problem. In: Desmedt JE, ed. New Developments in Electromyography and Clinical Neurophysiology. Basel, Switzerland: S Karger AG Medical and Scientific Publishers; 1973;1:519-532.

[7] Arsenault AB, Winter DA, Martiniuk RG. Is there a "normal" profile of EMG activity in gait? Med Biol Eng Comput. 1986;24:337-343.

[8] Kleissen RFM, Hermens HJ, den Exter T, et al. Simultaneous measurement of surface EMG and movements for clinical use. Med Biol Eng Comput. 1989;27:291-297.

[9] Yang J, Winter DA. Surface EMG profiles during different walking cadences in humans. Electroencephalogr Clin Neurophysiol. 1985;60:485-491.

[10] Humphreys DS. The Analysis, Design and Synthesis of Electrical Filters. Englewood Cliffs, NJ: Prentice-Hall Inc; 1970:38-45.

[11] Winters JM, Stark L. Muscle models: what is gained and what is lost by varying model complexity. Biol Cybern. 1987;55:403-420.

[12] Coggshal JC, Bekey GA. EMG-force dynamics in human skeletal muscle. Med Biol Eng. 1970;8:265-270.

[13] Crosby PA. Use of surface electromyogram as a measure of dynamic force in human limb muscles. Med Biol Eng Comput. 1978;16:519-524.

[14] Kleissen RFM, Hermens HJ. A muscle model relating hip abductor torque and surface EMG. In: Proceedings of the 12th International Congress of Biomechanics; 1989; Los Angeles, Calif.

[15] Winter DA, Yack HJ. EMG profiles during normal human walking: stride-to-stride control and inter-subject variability. Electroencephalogr Clin Neurophysiol. 1987;67:402-411.

[16] Ericson MO, Nisell R, Ekholm J. Quantified electromyography of lower limb muscles during level walking. Scand J Rehabil Med. 1986;18:159-163.

R Kleissen, MSc, is Research Engineer, Research and Development Group, Rehabilitation Centre Het Roessingh, Roessinghsbleekweg 33, 7522 AH Enschede, the Netherlands.

This work was supported by a grant from St Jorisstichting, Bussum, the Netherlands.