Sampling rate effects on surface EMG timing and amplitude measures

Clinical Biomechanics 18 (2003) 543–552 www.elsevier.com/locate/clinbiomech

Sampling rate effects on surface EMG timing and amplitude measures Jeffrey C. Ives *, Janet K. Wigglesworth Department of Exercise and Sport Sciences, Center for Health Sciences, Ithaca College, Ithaca, NY 14850, USA Received 18 October 2002; accepted 11 April 2003

Abstract Objective. To determine if oversampling the surface electromyographic signal provides any benefit in analyzing common electromyographic timing and amplitude measures used in kinesiological studies. Design. A within subjects (n ¼ 8) repeated measures design was used to examine surface electromyographic signals captured under four contraction modes and acquired with five different analog-to-digital sampling rates. Background. There is a growing trend to sample surface electromyography at rates higher than the Nyquist rate. Though there is limited evidence to support oversampling, the necessity or benefit of doing so remains unclear. Methods. Surface electromyography was recorded from the triceps brachii during maximal, submaximal, and fatiguing isometric contractions, as well as dynamic contractions. The analog signals were bandpassed between 20 Hz and 2 kHz, and oversampled at 6 kHz. The signals were then digitally resampled at 3 kHz, 1 kHz, 500 Hz, and 250 Hz without benefit of an anti-aliasing filter. Amplitude and timing variables measured from both the rectified and smoothed signal were compared across sampling rates. Results. Oversampling produced no significant changes in timing and amplitude measures of the rectified or smoothed electromyographic signal. For the smoothed signal, minor undersampling at half the Nyquist rate was sufficient to accurately capture most timing and signal strength measures. Conclusions. Oversampling is unnecessary to gather typical amplitude and timing measures from the surface electromyographic signal. Electromyography sampled below half the Nyquist rate is likely to result in a poor temporal and amplitude representation of the signal. Relevance Computer memory and processing resources for analyzing amplitude and timing information need not be expended in oversampling surface electromyography, and results of previous studies need not be outright dismissed because of minor undersampling violations or the lack of an anti-aliasing filter.
2003 Elsevier Science Ltd. All rights reserved. Keywords: Analog to digital conversion; Sampling rate; Surface EMG; Measurement

1. Introduction Surface electromyography (sEMG) is a common method to investigate muscle action in kinesiological studies and musculoskeletal rehabilitation. Capturing and storing the sEMG signal is most commonly done digitally by computer, which requires converting the analog signal to a digital signal by way of an analog to digital (A/D) converter. An important factor in A/D conversion is the sampling rate. Too slow a sampling

*

Corresponding author. E-mail address: [email protected] (J.C. Ives).

rate can result in distortion of the signal, such as aliasing (Gitter and Stovlov, 1995). To avoid aliasing and other signal distortion, the sampling rate must be a least double the highest frequency components in the signal (Clancy et al., 2002). This 2· rate is known as the Nyquist rate. The highest frequency components of the sEMG signal have been reported to be around 400–500 Hz (Clancy et al., 2002; Winter, 1990), thus convention recommends sampling rates of 800–1000 Hz, particularly when used with an anti-aliasing filter with a high frequency cutoff of 400–500 Hz (Hermens et al., 1999). Several authors have suggested that the Nyquist rate is too slow and sampling rates should be 3–10· the

0268-0033/03/$ - see front matter
2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0268-0033(03)00089-5

544

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

maximal frequency content in the signal (Gitter and Stovlov, 1995; Nilsson et al., 1993; Pagnacco et al., 1997). Sampling rates below 5· the maximal frequency content of the signal limits waveform resolution (Gitter and Stovlov, 1995), which results in errors when analyzing the EMG signal for waveform patterns like turns and spike amplitudes (Jørgensen and Fuglsang-Frederiksen, 1991). Surface EMG used in typical kinesiological applications does not require waveform analyses; yet even common timing and amplitude measures of the signal may not be accurately measured from a signal captured at the Nyquist rate. For example, Nilsson et al. (1993) reported ‘‘dramatically’’ altered onset timing of evoked muscle sEMG signals when sampled below 4·. Using a generated sinusoidal waveform, these authors also found marked decreases in signal amplitude when sampled below 3·, but these results were not consistent with the evoked muscle sEMG. Nonetheless, the common use of the 800–1000 Hz sampling rate for sEMG is put in question, even when using a 500 Hz anti-aliasing low-pass filter. These data also put in question the interpretation of past experiments using sampling rates equal to or below 1 kHz for sEMG––many of which did not use anti-aliasing filters. It is clear that slow sampling rates distort the frequency content and waveform shape of the sEMG signal, but there is little empirical evidence to show how, or how much, this distortion affects the two most common measures taken from kinesiological sEMG records; amplitude (signal strength) and timing. Put differently, does the sEMG signal oversampled above the Nyquist rate provide different information, and at what sampling rate is there a meaningful loss of information? Thus, the purpose of this investigation was to investigate the effects of sampling rate on timing and amplitude measures of the sEMG signal and determine if the Nyquist rate is sufficient for typical kinesiological uses.

2. Methods Because different types of muscle contractions produce different motor unit behaviors (Conwit et al., 1999), slow versus fast sampling rates may differentially effect sEMG signals recorded from contractions varying in force, speed of movement, or in fatigued versus nonfatigued muscle. Hence, sEMG signals were recorded from maximal isometric contractions, submaximal isometric contractions, rapid dynamic contractions, and fatigued muscle isometric contractions. In addition, both subjective (visually determined) and objective (criterion-based) measures of sEMG burst onsets and offsets were analyzed.

2.1. Subjects Eight apparently healthy volunteers (5 male and 3 female, mean age ¼ 24.5, age range ¼ 16–38 years) provided informed consent to participate according to institutional guidelines. A medical history form was completed by each subject to check for neuromuscular pathology and to clear them to perform vigorous contractions. 2.2. Procedures Testing was on one day only. Subjects sat in a hydraulic resistance dynamometer (UNEX II; Sammons– Preston, Bolingbrook, IL, USA) set up to assess elbow extension strength of the right arm. Subjects were positioned with their upper arm vertical and forearm horizontal for a relative elbow angle of 90. The posterior aspect of the upper arm was fixed against the chair backrest to restrict extraneous motion. The dynamometer lever arm was positioned with its axis of rotation aligned with the elbow axis. With their forearm pronated, the subjectsÕ gripped a handle on the lever arm. Following electrode preparation and subject familiarization, the subjects performed elbow extension contractions in the following order: (1) three maximal voluntary isometric contractions (Max), (2) three submaximal contractions at 50% max (Submax), (3) three fast as possible dynamic contractions (Dynamic), and (4) three max contractions following a fatiguing protocol (Fatigue). All contractions were initiated with an auditory stimulus to which subjects were instructed to react as fast as possible. Each isometric contraction lasted 3–4 s and stopped on verbal command from the experimenter. One minute separated each contraction and three minutes separated each condition. The target load for the submax condition was 50% of the subjectÕs maximal force determined from the highest force produced in the previous three maximal trials. The hydraulic gauge of the dynamometer provided visual force feedback to the subjects to give them a target force for the submax condition. For fast dynamic contractions, the subjects performed three maximal speed elbow extensions in the absence of any resistance other than the inertia of the lever arm. Subjects were instructed to extend to near full range of motion and rapidly return to the starting position which was located with a mechanical block to the lever arm. The fatigue protocol before the last set of max contractions consisted of rapidly repeating isokinetic maximal effort elbow extensions performed with the isokinetic resistance set equivalent to 50% of their maximal isometric strength. Subjects performed the contractions––normally 40–50 repetitions––until their

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

force dropped below 40% of their maximal isokinetic force. The three maximal isometric contraction trials immediately followed the fatigue protocol and no rest was provided between trials to avoid recovery. To gather evidence that the contraction modes resulted in different motor unit behaviors, median power frequencies (MPFs) were calculated from the 6 kHz sampling rate data with fast Fourier transformations using a Hamming window on de-meaned and de-trended time domain data (Kamen and Caldwell, 1996). MPFs were calculated for two trials from each contraction mode for all subjects. 2.3. Electromyography Before electrode placement, the skin was prepared according to standard procedures. Hair was shaved if necessary, the skin was cleansed with alcohol, and the skin was lightly abraded with an emery cloth. Two eight mm diameter Ag–AgCl passive electrodes (Sensormedics, Yorba Linda, CA, USA) were filled with electrode gel and placed in a bipolar configuration on lateral head of the right triceps brachii, just medial to the muscle belly in order to maximize distance from the biceps brachii. The electrodes were located parallel to the direction of the muscle fibers and placed 30 mm apart (center to center distance). A ground electrode was placed over the ipsilateral clavicle. Inter-electrode resistance was determined by an Ohmmeter and a level of less than 5000 X was accepted. The electrodes were connected to an EMG oscilloscope (Medic Flexline-S; Medical Instruments Co., Los Angeles, CA, USA) at a gain setting of 1000·, and bandpassed between 20 Hz and 2 kHz. The relatively high 2 kHz cutoff permitted capture of all signal frequencies so that sampling rate effects could be compared relative to the cutoff frequency of the amplifier (e.g., a 6 kHz sampling rate is 3· the maximum frequency permitted by the 2 kHz filter) and compared relative to the maximal sEMG frequencies producing any signal strength of consequence (e.g., a 6 kHz sampling rate is 12· the oft-reported 500 Hz maximum frequencies of the sEMG signal, or 6· the Nyquist rate). The 2 kHz cutoff, coupled with fast sampling rates, also has the undesired effect of capturing unwanted high frequency noise that originates primarily from the electronic equipment itself. Visual inspection of the baseline sEMG signals confirmed that quiet baseline activity–– relative to the amplitude of the desired sEMG signal–– was obtained for all subjects. The analog signal was routed through a 16 bit A/D system with 1.0 MX input impedance (Biopac Systems, Inc., Santa Barbara, CA, USA) connected to a Pentiumbased microcomputer. Data were sampled at 6 kHz, which again, oversampled by 3· the signal relative to the high frequency cutoff of the EMG amplifier and 12·

545

relative to the presumed maximal 500 Hz frequency in the sEMG signal. 2.4. Data analysis All data were analyzed offline with commercially available software (Acknowledge 3.3.2; Biopac Systems, Inc.). The software permitted the original 6 kHz digital data to be resampled without anti-aliasing modification at 3 kHz, 1 kHz, 500 Hz, and 250 Hz; a method used by others (Jayne et al., 1990) to examine sampling rate effects on non-mammalian muscle. This method was chosen over storing the analog sEMG data on analog FM tape and then playing back through the A/D system, because tape speed oscillations produces error and noise (Clancy et al., 2002). Upon playback these oscillations could introduce errors not associated with, but hard to distinguish from, sampling rate differences. At the sampled frequency and at each resampled frequency the following procedures and measures were taken: 1. Digitally sampled signals were full-wave rectified. From the rectified sEMG signal the following variables were measured: a. Visually determined onset time (Onset-VD). The onset point was determined subjectively by visual inspection of the first major vertical deflection of the sEMG signal (Gottlieb et al., 1989). Onset time was calculated as the time from the reaction stimulus to the onset of the sEMG burst. b. Visually determined burst duration. The offset point was determined subjectively by visual inspection of the last major vertical deflection of the sEMG signal. Burst duration was calculated as the difference between the onset and offset points. 2. The point of peak sEMG of the rectified signal was found by the software and the following variables determined: a. Timing of the peak, determined as the time from the reaction stimulus to the peak. b. Peak amplitude of the sEMG signal. 3. Data were then smoothed by a moving average using a window length of 5% of the sampling rate (e.g., 6 kHz sampling rate used a window of 300 data points, equivalent to a 50 ms window). From the smoothed sEMG signal the following variables were measured: a. Onset time of the smoothed signal (Onset10%max). The onset point was measured objectively at the point where the signal met or exceeded 10% of the peak signal amplitude (see Appendix A for more discussion). Onset-10%max was calculated as time from the reaction stimulus to this onset point. b. Burst duration of the smoothed signal. The offset point was calculated as the point where the signal

546

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

met or went below 10% of the peak signal amplitude. Burst duration of the smoothed signal was then calculated as the difference between the onset and offset points. c. Timing of the peak of the smoothed sEMG signal was measured. Time base of the peak was determined from the reaction stimulus to the peak. d. Peak amplitude of the smoothed sEMG signal. e. Average amplitude of the entire smoothed sEMG burst was measured. Burst duration was determined by the 10%max burst duration method. f. Total electrical activity (area under the curve) of the entire smoothed sEMG burst. Burst duration determined by the 10%max burst duration method. 4. The original 6 kHz sEMG trace was reloaded into the analysis program and full wave rectified in order to measure the onset point in a third manner: a. The onset point of the rectified signal was found using a two standard deviation threshold criteria (Onset-SD). The point was determined objectively as the point where the signal amplitude rose above the baseline mean value plus two standard deviations, and maintained at or above that level for at least 20 ms (see Appendix A for more details). Following the analysis of one trial the same trial was reloaded into the analysis program and resampled at the next slower sampling rate. These procedures were repeated by the same investigator until all trials were analyzed. The majority of the data were analyzed using a twoway repeated measures A N O V A (sampling rate · contraction type). Trial effects were not relevant to this investigation, so data from the three trials for each condition were averaged. Contrasts analyses were used to determine at what point, if any, a change in sampling rate led to significant changes in sEMG amplitude or timing. A separate three-way repeated measures A N O V A (measurement method · sampling rate · contraction type) was used to compare the three onset time measurement methods (Onset-VD, Onset-SD, Onset-10%max). A reliability analysis was performed to assess the degree of error in determining the onset point using the subjective visual inspection technique (Onset-VD). Only the 6 kHz sampling rate was chosen because this condition is most prone to variability. The 6 kHz rate was analyzed before the other sampling rates, and therefore the researcher was not influenced by prior data inspection. After all trials were analyzed by one researcher, a second researcher analyzed two of the three trials in each contraction mode, and repeated this on a second day. The second researcher was blinded to first researcherÕs data, and blinded on the second day to the previous dayÕs data. Intra-class correlation was used to assess inter- and intra-investigator consistency. For the final data set only the data assessed by the first re-

searcher was used. A significance level of 0.05 was used for all analyses, and the Hunyh–Feldt adjusted P -value was used in all cases.

3. Results 3.1. Reliability There was no significant difference between the Onset-VD measured by the first investigator (117.4 ms, SD 47.9) and the second investigator (114.8 ms, SD 47.5). An inter-rater correlation of 0.98 found. The mean Onset-VD assessed by a single investigator on the first day (114.8 ms, SD 47.5) did not significantly differ from that assessed on the second day (115.0 ms, SD 47.8). For a single rater the standard error of measurement for a single trial was 6.25 ms (95% confidence level) and an intra-rater intra-class correlation coefficient of 0.99. These results indicate highly reliable measurement techniques. 3.2. Contraction mode differences MPFs significantly (P ¼ 0:008) differed across the Max (107.4 Hz, SD 17.6), Submax (113.6 Hz, SD 24.1), Dynamic (115.7 Hz, SD 29.1), and Fatigue (86.9 Hz, SD 15.3) conditions. All timing and amplitude variables, except onset time, displayed significant differences across contraction modes. These results substantiate that the different contraction modes required different motor behaviors and contractile intensities. 3.3. Onset time measurement methods and sampling rate effects Fig. 1 illustrates a representative trial and Table 1 shows means and standard deviations for the onset time variables gathered from the rectified sEMG and the smoothed sEMG signal. Inspection of the mean values suggests that the Submax condition resulted in longer onsets, but this effect was not significant (P ¼ 0:14). The three-way repeated measures A N O V A revealed a significant (P ¼ 0:0001) measurement method  sampling rate interaction when comparing the three onset time measurement methods. Further analyses indicated that this interaction, illustrated in Fig. 2, was due to two things. First, Onset-10%max was significantly (P ¼ 0:0001) longer than Onset-SD, and this effect was consistent across sampling rates. Second, Onset-VD significantly (P ¼ 0:003) decreased with slower sampling rates while Onset-SD and Onset-10%max were not significantly affected by sampling rate. Onset-VD was analyzed separately to locate the source of the significant sampling rate effects. Despite a significant linear effect, Helmhert contrasts analyses indicated

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

547

nation for the high Onset-SD standard deviations could be non-systematic errors in onset determination observed in 9% of the trials as the sampling rate changed. Fig. 3 illustrates this phenomenon, where the Onset-SD is the same for 6 and 1 kHz, but at 3 kHz the waveform changed in such a way as to cause a false positive in the detection criteria. 3.4. Sampling rate effects on other timing variables

Fig. 1. Representative trial from a maximal voluntary isometric contraction sampled at 6 kHz and resampled at 1 kHz and 250 Hz (resampling at 3 kHz and 500 Hz not shown). Smoothed sEMG signal is superimposed on the rectified signal for illustrative purposes only. The 1 kHz signals look nearly identical to the 6 kHz signals. The 250 Hz signals look normal, but differ in many ways to the 6 and 1 kHz signals.

that the 2–6 ms differences between adjacent sampling rates were not significantly different from one another; significance was only found when comparing the highest sampling rates (6, 3, 1 kHz) to the slow 250 Hz rate. There were additional observations associated with the onset detection methods that were not readily teased out using statistical procedures, but have practical importance. The standard deviation of Onset-SD calculated across all sampling rate and contraction mode conditions (114.4 ms, SD 39.8) was larger than the standard deviation of Onset-10%max (120.7 ms, SD 30.7), and more like the standard deviation of the subjective Onset-VD (110.7 ms, SD 38.2). A partial expla-

Visually determined burst duration measured on the rectified signal significantly lengthened nearly 39 ms from the 6 kHz rate to the 250 Hz rate (Table 2). As with the rectified signal onset time, post hoc contrasts indicated no significant differences between adjacent sampling rates; significant effects were found only when comparing the fastest sampling rates (6, 3, 1 kHz) to the slowest rates (500, 250 Hz). Specifically, 6 and 3 kHz differed from 500 and 250 Hz, whereas 1 kHz differed only from 250 Hz. The 17 ms difference between the 1 kHz Nyquist rate and the 6 kHz oversampling was not significantly different. Burst duration measured on the smoothed signal was not significantly affected by sampling rate. Time to peak sEMG of the rectified signal and the smoothed signal were not significantly affected by sampling rate, nor were there significant contraction mode  sampling rate interactions. These results are somewhat misleading to the real effects of sampling rate on the sEMG pattern. Visual inspection of the sEMG signal revealed that the peak sometimes changed location when sampled at 500 or 250 Hz. For example, Fig. 4 illustrates that at 6 and 1 kHz sampling rates the peak sEMG of the rectified signal was at the second of two distinct peaks, whereas at 500 Hz the first peak was captured as the peak of the burst. This changing location of the peaks was not evident in the smoothed signal. Figs. 3 and 4 also illustrate that the low-level activity of the rectified signal is different at 1 kHz and slower sampling rates when compared to 6 kHz. 3.5. Sampling rate effects on amplitude variables Peak sEMG of the rectified signal significantly decreased with slower sampling rates (Table 3). Contrasts analyses revealed that the 99 lV (2.4%) drop in amplitude from 6000 through 1000 sampling rates was not significant, but the further drop in amplitude of 5.8% (500 Hz) and 11.4% (250 Hz) was significant. Peak amplitude of the smoothed sEMG signal was confounded by a significant sampling rate · contraction mode effect. From contrasts analyses it was found that peak amplitude of the smoothed sEMG signal did not significantly differ from 6 kHz to 500 Hz, but at 250 Hz the peak amplitude increased for the isometric contractions and decreased for the rapid dynamic contraction.

548

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

Table 1 Onset time means (SD) across contraction mode and sampling rate Contraction mode

Sampling rate (Hz) 6000

3000

1000

Onset-VD (ms)a Max Submax Dynamic Fatigue

103.8 140.8 115.8 108.6

(22.2) (64.4) (29.6) (36.1)

101.3 135.3 114.2 108.6

Onset-SD (ms)a Max Submax Dynamic Fatigue

100.1 133.5 112.6 106.5

(28.4) (61.8) (33.0) (39.4)

99.7 (28.6) 133.4 (60.7) 113.0 (32.9) 106.8 (39.4)

Onset-10%max (ms)a Max Submax Dynamic Fatigue

110.3 132.2 111.9 122.7

(19.8) (40.3) (29.6) (32.1)

110.4 133.7 112.2 122.1

(24.2) (58.6) (33.8) (38.9)

(19.7) (41.3) (29.3) (32.5)

500

250

99.8 (24.3) 123.4 (47.8) 105.5 (37.3) 100.3 (38.4)

96.8b (20.8) 116.3b (47.1) 96.7b (29.2) 98.2b (30.3)

99.6 (28.7) 135.5 (61.0) 114.3 (32.1) 107.0 (39.0)

104.5 133.4 114.1 107.9

(26.4) (53.2) (31.4) (34.3)

103.8 132.2 116.5 112.2

(24.5) (48.6) (31.5) (38.0)

110.5 133.2 112.5 123.8

114.5 133.7 113.7 123.6

(23.6) (38.9) (30.2) (30.2)

117.1 136.6 112.3 126.6

(22.7) (39.8) (28.3) (31.1)

100.2 132.7 111.3 105.2

(26.1) (52.4) (32.6) (35.6)

(19.4) (40.0) (29.3) (31.9)

a Onset-SD and Onset-10%max significantly differ from one another. Both significantly differ from Onset-VD, but these differences are dependent on sampling rate. b 250 Hz sampling rate significantly different from 6, 3, and 1 kHz.

130

Onset Time (ms)

125 120 115 110 105

onset-VD

100

onset-SD onset-10%max

95 6000

3000

1000

500

250

Sampling Rate (Hz)

Fig. 2. Plot of the sampling rate  measurement method interaction for onset times. All onset methods significantly differ among each other at each sampling rate except for the three comparisons that are circled. Error bars are ±SE.

Average sEMG of the rectified signal was also confounded by a significant sampling rate · contraction mode interaction. Average sEMG decreased with slower sampling rates, but only for the dynamic contraction. In the dynamic contractions the average rectified sEMG amplitude significantly fell 43 lV (7.4%) from 6 kHz to 500 Hz, and 111.3 lV (19.2%) from 6 kHz to 250 Hz. On the other hand, average sEMG and total sEMG of the smoothed signal were not significantly affected by the different sampling rates. 4. Discussion This investigation found that oversampling above the Nyquist rate has statistically non-significant effects on

timing and amplitude measures of the surface EMG signal. Undersampling at 500 and 250 Hz had statistically significant effects on onset latency and burst duration measures when using visual determination criteria of the rectified signal. Peak amplitude measures of the rectified signal systematically decreased with slower sampling rates, but again, this effect was only significant at 500 and 250 Hz. Figs. 1 and 4 illustrate that at the undersampled rates the rectified waveform shape may change to such an extent that the appearance and timing of the peaks is dramatically altered. Smoothing the sEMG signal counteracted most of the negative effects of undersampling. Specifically, burst onset and duration of the smoothed signal, and average sEMG amplitude and total sEMG area, were not significantly affected even when sampled at 250 Hz. Though the peak sEMG amplitude of the smoothed signal was not significantly different when undersampled at 500 Hz, the location of the peak value may change in the rectified signal. If the location of the peak changes in the rectified signal the potential for the peak to change in the smoothed signal exists. These findings confirm the conventional 1 kHz rate as the Nyquist sampling rate for sEMG and that sampling over this is unnecessary to accurately capture fundamental timing and amplitude measures of the sEMG signal. A 1 kHz sampling rate is sufficient to capture important properties of sEMG signal even without an anti-aliasing filter, that is, fundamental timing and amplitude measures are not affected by aliasing that may occur to frequency components above 500 Hz. Though it is not recommended that data be undersampled, the present results suggest that undersampling at 500 Hz does not automatically invalidate timing or amplitude

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

549

3.5

Amplitude (mV)

3.0

6000 Hz

2.5 2.0 1.5 1.0 0.5 0.0

Amplitude (mV)

3.5 3.0

1000 Hz

2.5 2.0 1.5 1.0 0.5 0.0 3.5

Amplitude (mV)

3.0

500 Hz

2.5 2.0 1.5 1.0 0.5 0.0 50

Fig. 3. Representative trial illustrating a non-systematic error in OnsetSD at high sampling rates. All traces have been clipped in order to magnify baseline activity. The onset points (arrows) at 6 and 1 kHz were selected at about the same time. The onset at 3 kHz was selected at a much earlier point despite using the same threshold measurement criteria.

data if the sEMG signal has been appropriately smoothed. These findings are contrary to suggestions that oversampling up to 5–10· may be necessary to gather valid data (Gitter and Stovlov, 1995; Nilsson et al., 1993). Using a generated sinusoidal waveform and varying the ratio of sampling rate to waveform frequency, Nilsson et al. (1993) found that when the sampling rate fell below 4 the upper frequency of the signal that amplitude errors in the range of 10% occur, and at the 2· (Nyquist) rate amplitude errors of 35% occur. Collecting evoked muscle action potentials with surface EMG, however, peak to peak amplitude data reported by these authors were similar to that of the rectified data in the current study, that is, slight drops in amplitude at the Nyquist rate compared to oversampling, and a larger (40%) drop when undersampled. Others have reported no systematic changes in the peak or total EMG amplitudes sampled at or above the Nyquist rate (Jayne et al., 1990; Sadhukhan et al., 1994).

100

150

200 ms

Fig. 4. Representative trial from a rapid dynamic contraction sampled at 6 kHz and resampled at 1 kHz and 500 Hz. Smoothed sEMG signal is superimposed on the rectified signal for illustrative purposes only. At 500 Hz the peak amplitude of the rectified signal changes location. Note the marked similarities of the smoothed signal across sampling rates.

There is limited evidence describing how sampling rate affects surface EMG burst timing. Nilsson et al. (1993) noticed large changes in onset latencies and burst durations of evoked surface EMG signals that were not sampled above the Nyquist rate, but a lack of numerical data and unconvincing illustrations cast doubt on their report. The present data indicate that sEMG burst timing is not significantly affected by oversampling, and undersampling does not affect all timing detection methods equally. Undersampling affected visually determined timing measures of the rectified signal and did not significantly influence measures of burst onset or duration when measured using objective threshold criteria. Appropriately set objective criteria inherently take into account waveform differences due to sampling rate. The observation that the Onset-SD method produced errors at the 3, 1 kHz, and 500 Hz sampling rates were too few and random to contradict this suggestion, but rather, indicate what can happen when a highly sensitive threshold criteria is used without confirmation by visual inspection.

550

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

Table 2 Burst duration and time to peak sEMG means (SD) across contraction mode and sampling rate Contraction mode

Sampling rate (Hz) 6000

3000

1000

500

250

2151.0 (143.7) 2343.0 (153.8) 326.9 (200.5) 2066.6 (115.8)

2150.5 (147.4) 2353.9 (151.0) 328.8 (208.3) 2068.0 (130.7)

2159.2 (157.9) 2364.2 (141.7) 345.9 (226.0) 2085.3 (121.5)

2184.0a (138.5) 2386.7a (135.8) 363.7a (249.2) 2093.1a (117.5)

2177.4b (162.8) 2386.5b (127.2) 395.3b (245.7) 2078.3b (121.6)

2116.0 (125.1) 2342.0 (124.6) 352.5 (206.4) 2036.8 (131.2)

2114.0 (121.8) 2350.7 (125.8) 343.4 (212.8) 2038.5 (133.5)

2114.9 (123.2) 2317.2 (131.5) 346.6 (216.9) 2046.6 (127.5)

2122.2 (120.9) 2343.2 (103.0) 318.6 (212.0) 2049.9 (141.1)

2118.7 (128.0) 2365.2 (115.0) 317.3 (188.4) 2040.0 (128.4)

Time to peak EMG-rectified signal (ms) Max 1103.2 (352.6) Submax 1396.4 (389.3) Dynamic 236.8 (118.5) Fatigue 1080.5 (222.5)

1106.1 (353.7) 1359.4 (356.3) 237.0 (119.0) 1080.4 (222.9)

1055.0 (434.6) 1325.8 (327.7) 236.0 (117.7) 1082.6 (223.7)

993.2 (385.8) 1346.5 (402.7) 230.2 (112.7) 1116.7 (195.6)

1135.0 (288.1) 1353.0 (511.5) 244.0 (127.8) 1074.3 (286.8)

Time to peak EMG-smoothed signal (ms) Max 1158.8 (343.7) Submax 1200.5 (409.6) Dynamic 210.2 (78.9) Fatigue 1139.5 (304.6)

1144.1 (341.3) 1198.8 (408.5) 214.0 (76.6) 1157.2 (339.6)

1180.3 (378.4) 1200.9 (409.3) 210.8 (79.3) 1118.6 (297.3)

1188.7 (368.0) 1160.7 (385.4) 212.5 (80.2) 1178.7 (306.8)

1083.7 (350.5) 1139.4 (473.4) 215.8 (79.3) 1179.7 (298.3)

Burst duration-rectified signal (ms) Max Submax Dynamic Fatigue Burst duration-smoothed signal (ms) Max Submax Dynamic Fatigue

a b

500 Hz sampling rate significantly different from 6 and 3 kHz. 250 Hz sampling rate significantly different from 6, 3, and 1 kHz.

Visual inspection was markedly influenced by undersampling. Visual pattern recognition of the rectified signal may have been altered as waveform patterns became simplified with slower sampling rates. The loss of resolution in the undersampled signal may have contributed to these differences. Clearly, undersampling should not be used to facilitate visual inspection. In conclusion, these data provide no evidence that oversampling sEMG above the Nyquist rate in typical kinesiological investigations provides any benefit. A 1 kHz sampling rate can be considered the functional Nyquist rate even without an anti-aliasing filter. Based on the findings here it may be tempting to undersample and smooth the sEMG data, but this is not advised. All sEMG data should be saved in raw form (Soderberg and Knutson, 2000), and if these data are undersampled

their usefulness is limited. In addition, although the small differences among sampling rates reported here were non-significant, these differences could be important in some circumstances. For example, the 17 ms difference in Onset-VD between the 6 kHz and 500 Hz sampling rates in the submax contraction probably holds no relevance in gait analysis, but could be quite meaningful in studies of reaction time. Clinical relevance, apart from statistical significance, must be considered. Further study is warranted to examine sampling rate effects with different methods of EMG processing (e.g., root mean square), different electrode sizes and configurations, different objective threshold criteria, and with different muscles. In addition, it would be useful to examine the use of an anti-aliasing filter to determine not only if these results remain valid, but also if undersampling is a viable option.

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

551

Table 3 Amplitude variable means (SD) across contraction mode and sampling Contraction mode Peak EMG-rectified signal (lV) Max Submax Dynamic Fatigue Peak EMG-smoothed signal (lV) Max Submax Dynamic Fatigue Average EMG-rectified signal (lV) Max Submax Dynamic Fatigue Average EMG-smoothed signal (lV) Max Submax Dynamic Fatigue Total EMG-smoothed signal (lV ms) Max Submax Dynamic Fatigue a b

Sampling rate (Hz) 6000

3000

1000

500

250

5005.0 (2144.9) 3109.4 (1747.5) 3370.2 (1321.1) 5142.5 (2461.9)

4995.9 (2142.3) 3100.0 (1746.9) 3354.4 (1322.0) 5126.3 (2454.8)

4893.2 (2139.5) 2993.8 (1685.3) 3278.2 (1360.3) 5064.3 (2444.8)

4785.4a (2151.1) 2844.8a (1644.9) 3021.0a (1151.8) 5004.3a (2429.1)

4519.1a (2130.5) 2673.7a (1600.1) 2780.4a (1092.3) 4765.9a (2393.3)

1671.3 (868.0) 985.1 (561.6) 1169.2 (454.6) 1879.8 (919.9)

1670.5 (867.8) 988.9 (558.1) 1167.8 (454.4) 1865.8 (904.2)

1662.5 (865.8) 978.4 (559.8) 1163.0 (450.1) 1870.3 (912.6)

1654.2 (851.3) 992.7 (572.5) 1196.3 (415.4) 1891.9 (924.5)

1765.7b (915.7) 1006.7b (576.5) 1151.4b (465.7) 1931.3b (952.2)

939.0 (517.6) 459.3 (259.1) 579.0 (212.4) 1002.1 (517.0)

939.1 (517.3) 459.3 (261.8) 553.4 (210.4) 1001.8 (515.4)

930.3 (510.8) 457.0 (258.9) 551.3 (214.0) 994.8 (513.5)

924.7 (508.0) 452.0 (254.4) 535.9a (224.5) 996.0 (515.4)

925.3 (511.5) 452.2 (257.7) 467.7a (213.2) 1000.4 (517.1)

887.0 (487.8) 440.0 (250.8) 404.1 (160.0) 954.4 (498.0)

883.8 (492.0) 435.9 (245.9) 379.2 (134.5) 941.0 (492.3)

880.0 (494.3) 431.3 (244.3) 383.7 (145.0) 940.5 (491.3)

880.3 (486.1) 431.8 (245.9) 384.8 (136.7) 948.0 (497.9)

881.2 (492.5) 443.3 (240.1) 385.4 (151.4) 951.8 (502.3)

1998.1 (1223.3) 1041.7 (659.9) 148.2 (75.0) 2024.2 (1131.2)

2009.2 (1219.2) 1043.0 (659.7) 156.1 (80.8) 2027.4 (1132.2)

2008.2 (1219.1) 1037.8 (659.1) 149.2 (77.1) 2028.8 (1134.2)

1963.1 (1158.0) 1043.8 (651.5) 150.0 (77.2) 2030.7 (1124.8)

2000.1 (1220.9) 1032.0 (662.7) 146.4 (77.9) 2026.5 (1133.1)

500 and 250 Hz significantly different from 6, 3, and 1 kHz. 250 Hz significantly different from 6, 3, and 1 kHz, but effect differs depending on contraction mode.

Acknowledgements

Appendix A

The authors deeply thank Celeste Gabai for her dedicated and skillful work in the initial digitization of the data set, and Joshua McCaig for his data collection assistance. This study was supported by a grant from the School of Health Sciences and Human Performance at Ithaca College.

Computer-assisted criteria for determining EMG onsets and offsets typically establish a threshold level that is a percentage of the maximum signal amplitude, or is some number of standard deviations above the baseline mean. Both of these methods were used because of their simplicity and widespread use in the clinical

552

J.C. Ives, J.K. Wigglesworth / Clinical Biomechanics 18 (2003) 543–552

literature (e.g., Bullock-Saxton, 1994). There are more complex mathematical techniques (e.g., Micera et al., 1998), but these methods are cumbersome for rapid clinical decision making. Furthermore, there is little agreement on the ‘‘best’’ method of automatic or objective determination of onset and offset points (Hodges and Bui, 1996; Soderberg and Knutson, 2000). For simple and stable contractions with large amplitudes relative to baseline activity, the methods used here for determining threshold criteria have been deemed sufficient (for discussion, see Soderberg and Knutson, 2000; Staude and Wolf, 1999). In determining the specific threshold values and time windows as described below, trial and error experimentation was done to find values that would work well regardless of contraction mode and sampling rate. It could be that a particular threshold could work best for one type of contraction and sampling rate, but these data were non-existent. The two standard deviation threshold. The sEMG mean amplitude and standard deviation were recorded during a 50 ms epoch (52 ms for the 250 Hz sampling rate) of baseline activity occurring during the first 100 ms of the sEMG trace. Generally taken from 20 to 70 ms, the 50 ms epoch was visually inspected to insure that no obvious extraneous contractile activity or noise were present. The sEMG mean plus two standard deviations was used as an amplitude threshold to determine the onset point of the sEMG burst. Using a sliding window of 20 ms duration, the first point at which the amplitude rose above the criterion threshold and remained at or above the threshold for 20 ms was selected (Hodges and Bui, 1996). In order to minimize subjectivity, this point was selected regardless of how it compared to the visually determined onset point. The 6 kHz sampling rate trace was analyzed first, followed by the resampled traces in descending order. The same time epoch was used for each resampled trace. In two trials the subject anticipated the go signal and reacted well before 100 ms, leaving no appropriate epoch to determine baseline activity before the start of the contraction. In these two trials the 50 ms baseline epoch was found after the contraction had ended. The 10% of maximum criteria. Following smoothing of the sEMG trace, the computer automatically found the maximum value of the signal and calculated 10% of the maximum as the onset and offset criterion. The 10% value determining the onset point was found by moving the computer cursor along the apparent start of the sEMG burst. The first point at which the signal rose above the 10% criteria and stayed that way for at least 5 ms was selected. The same procedures were used to find the offset point, only the first point at which the signal fell below the 10% criteria at the apparent end of the burst was selected.

For the standard deviation method and the 10% of maximum method the time point was selected at the first point that met or exceeded the criterion threshold, resulting in some points not corresponding exactly with the threshold value. This error is due to the sampling rate resolution, which at 1 kHz and faster is 1 ms or less. At 250 Hz sampling rate, however, this error could be as large as 4 ms.

References Bullock-Saxton, J.E., 1994. Local sensation changes and altered hip muscle function following severe ankle sprains. Phys. Ther. 74, 17– 31. Clancy, E.A., Morin, E.L., Merletti, R., 2002. Sampling, noisereduction and amplitude estimation issues in surface electromyography. J. Electromyogr. Kines. 12, 1–16. Conwit, R.A., Stashuk, D., Tracy, B., McHugh, M., Brown, W.F., Metter, E.J., 1999. The relationship of motor unit size, firing rate and force. Clin. Neurophysiol. 110, 1270–1275. Gitter, A.J., Stovlov, W.C., 1995. AAEM minimonograph #16: Instrumentation and measurement in electrodiagnostic medicine––part II. Muscle Nerve 18, 812–824. Gottlieb, G.L., Corcos, D.M., Agarwal, G.C., 1989. Organizing principles for single-joint movements. I. A speed-insensitive strategy. J. Neurophysiol. 62, 342–357. Hermens, H., Freriks, B., Merletti, R., Stegeman, D., Blok, J., et al., 1999. European recommendations for surface electromyography. [SENIAM 8]. Roessingh Research and Development, Enschede, NL. Hodges, P.W., Bui, B.H., 1996. A comparison of computer-based methods for the determination of onset of muscle contraction using electromyography. Electroenceph. Clin. Neurophysiol. 101, 511– 519. Jayne, B.C., Lauder, G.V., Reily, S.M., Wainwright, P.C., 1990. The effect of sampling rate on the analysis of digital electromyograms from vertebrate muscle. J. Exp. Biol. 154, 557–565. ., Fuglsang-Frederiksen, A., 1991. Turns-amplitude Jørgensen, S.A analysis at different sampling frequencies. Electroenceph. Clin. Neurophysiol. 81, 1–7. Kamen, G., Caldwell, G.E., 1996. Physiology and interpretation of the electromyogram. J Clin. Neurophysiol. 13, 366–384. Micera, S., Sabatini, A.M., Dario, P., 1998. An algorithm for detecting the onset of muscle contraction by EMG signal processing. Med. Eng. Phys. 20, 211–215. Nilsson, J., Panizza, M., Hallett, M., 1993. Principles of digital sampling of a physiologic signal. Electroenceph. Clin. Neurophysiol. 89, 349–358. Pagnacco, G., Oggero, E., Morr, D.R., Berme, N., 1997. Oversampling data acquisition to improve resolution of digitized signals. Biomed. Sci. Instrum. 34, 137–142. Sadhukhan, A.K., Goswami, A., Kumar, A., Gupta, S., 1994. Effects of sampling frequency on EMG power spectral characteristics. Electromyogr. Clin. Neurophysiol. 34, 159–163. Soderberg, G.L., Knutson, L.M., 2000. A guide for use and interpretation of kinesiologic electromyographic data. Phys. Ther. 80, 485– 498. Staude, G., Wolf, W., 1999. Objective motor response onset detection in surface myoelectric signals. Med. Eng. Phys. 21, 449–467. Winter, D.A., 1990. Biomechanics and Motor Control of Human Movement. Wiley Interscience, New York.