Surface electromyography (EMG) is a very commonly used method for examining skeletal muscle fatigue. In many cases, an amplitude and center frequency (mean frequency or median frequency) parameter is used to try to determine the underlying cause(s) for the fatigue. As acknowledged by Basmajian and De Luca (1), an important consideration is the fact that fatigue is a process, rather than an end point. Thus, measurement of force and power does not always reflect the fatigue-related changes that occur inside the muscle, especially during a submaximal activity when force and power remain constant but the performance capabilities of the muscle decrease. One of the most consistent findings in early EMG studies was that during a fatiguing isometric muscle action, the bandwidth of the signal's power spectrum decreased, and the peak shifted to lower frequencies (5,10). Lindström et al. (10) determined that this response was mostly because of a decrease in muscle fiber action potential conduction velocity. Since these early investigations, a great deal of research has been conducted to further examine the EMG amplitude and center frequency responses during fatiguing activities. However, there are still many research questions that need to be answered, particularly in the area of fatiguing dynamic muscle actions.
Although dynamic muscle actions are much more common in athletics than are isometric muscle actions, they present some challenges when processing the resulting EMG signals. The most important challenge results from the fact that the signal can be nonstationary because of movement of the muscle fibers beneath the recording electrodes, changes in fiber length, and alterations in the number of active motor units and their firing rates (2). Thus, it has been suggested (2,9,15) that joint time-frequency signal processing techniques should be used to analyze EMG signals recorded during dynamic muscle actions because these methods are capable of tracking the changes in power that can occur across both time and frequency. There are 3 primary joint time-frequency methods: (a) the Short-Time Fourier Transform (4), (b) the Wigner-Ville Transform (7), and (c) the Wavelet Transform (15). Each of these techniques has advantages and disadvantages. For example, the Short-Time Fourier Transform is a very simple modification of the Discrete Fourier Transform, but its results are greatly influenced by the window length used for the analysis (i.e., short windows provide good time resolution but poor frequency resolution, and long windows provide poor time resolution but good frequency resolution). Unfortunately, there is no single window length that is optimal for all surface EMG signals, because some signals demonstrate very rapid changes in power and frequency, whereas others change more slowly. This makes it nearly impossible for the researcher to be confident that he or she is getting the best possible combination of time and frequency resolution for every EMG signal. The Wigner-Ville Transform and its variations have also been used to process EMG signals from dynamic muscle actions (3). An advantage of the Wigner-Ville Transform is that it has the best joint time-frequency resolution of the 3 methods. However, this resolution comes at a cost in the form of interference terms (aka cross terms or ghost terms) that can interfere with the primary signal terms. As described by Rioul and Vetterli (12), this is particularly problematic with multicomponent signals. An alternative to the Short-Time Fourier Transform and Wigner-Ville Transform is the Wavelet Transform, which decomposes the signal onto a set of localized (in both time and frequency) basis functions. The primary disadvantage of the wavelet transform is that it is a time-scale analysis, rather than a time-frequency analysis. Thus, linear changes in scale result in logarithmic changes in frequency. However, von Tscharner (15) has adapted the wavelet transform by using a filter bank of nonlinearly scaled Cauchy wavelets that provide the best possible combination of time and frequency resolution for EMG signals. These wavelets are capable of accurately identifying the intensities that comprise the EMG signal and localizing them in both time and frequency.
The result of the wavelet analysis developed by von Tscharner (15) is a time-frequency distribution that shows the changes in intensity across time at different frequencies, thereby reflecting the nonstationary characteristics of the signal. Because the intensity pattern is a 2-dimensional matrix of data, it cannot be analyzed with traditional univariate methods, such as parametric statistics. Recent studies (16,19,20), however, have begun using multivariate procedures, such as principal components analysis, to statistically analyze EMG intensity patterns. Although these methods have traditionally been applied to EMG signals for the purpose of controlling externally powered prostheses (6), they can also be used to extract detailed information from the EMG signal. Generally speaking, these multivariate procedures can be classified into 2 categories: (a) techniques that classify patterns into separate groups and (b) methods that track changes in patterns over time. In both cases, however, the intensity patterns must first be decomposed into a multidimensional pattern space (16). For example, von Tscharner and Goepfert (19) used pattern classification to determine if the EMG intensity patterns from the thigh and lower leg muscles during running could be accurately classified into respective gender-specific categories (men vs. women). The results showed that the patterns were accurately discriminated in >95% of the cases, and it was concluded that the overall pattern of muscle activation during running may be different for men and women. Pattern analysis methods were also used by von Tscharner (16) to examine the changes in EMG intensity patterns over time during low-intensity cycling. Instead of classifying the patterns into separate categories, however, they were tracked with a trend plot (16). The purpose of the trend plot is to statistically analyze the changes or trends in EMG intensity patterns that occur over time. In theory, when EMG intensity patterns change over time in a predictable manner, they move in pattern space along a path that can be tracked with linear regression and quantified with the trend plot. The results showed that the trend plot was capable of quantifying the complex changes that occurred in the EMG intensity pattern with mild fatigue (16).
The intensity patterns produced from the wavelet analysis are rich with information (15), but methods to analyze them are still being developed. Since the von Tscharner (16) study, no further investigations have used the trend plot to examine muscle fatigue. We feel that the principal components analysis and trend plot could be a very useful tool for researchers and practitioners that want to use the EMG signal to study muscle fatigue, particularly during dynamic muscle actions, when the signal is likely to be nonstationary. The trend plot may also reflect the differences between subjects in the percent decline in force and power that occurs when a muscle or muscle group becomes fatigued. If this is true, it could potentially be used by coaches and athletes to examine the fatigue status of individual muscles that cross the same joint. However, before these questions can be answered, more research needs to be performed with various types of fatiguing activities to evaluate the performance of the trend plot. Thus, the purpose of this study was to use the wavelet analysis developed by von Tscharner (15) combined with the trend plot to examine changes in EMG intensity patterns during maximal, fatiguing isokinetic muscle actions. We hypothesize that the trend plot will show systematic changes to the EMG intensity patterns with muscle fatigue. These changes will be consistent within subjects, but not necessarily between subjects, thereby reflecting subject-specific strategies for coping with fatigue.
Experimental Approach to the Problem
This study was designed to use an EMG-based tool called the trend plot to examine the complex changes that occur in muscle activation patterns during fatigue. Each subject performed 50 consecutive maximal concentric isokinetic muscle actions of the leg extensors as EMG signals were detected from the vastus lateralis. The EMG signals were processed with a wavelet analysis, which is capable of tracking changes in the intensity (or power) of the signal both in time and in frequency. The result is an EMG intensity pattern that is further analyzed with principal components analysis and the trend plot. If the EMG signals change consistently because of muscle fatigue, then these changes will be reflected in the trend plot. In many cases, these changes are very subtle and cannot be evaluated by the analysis of a single amplitude or frequency parameter. Thus, when combined with the trend plot, the wavelet analysis is a very powerful tool that can be used by coaches and athletes to track muscle fatigue. This tool could potentially be very useful in situations where the declines in force or power output of individual muscles with fatigue cannot be monitored (e.g., with the individual quadriceps femoris muscles).
Eleven men (mean ± SD age = 22.4 ± 1.1 years; height = 178.9 ± 5.3 cm; body weight = 79.6 ± 10.6 kg) and 7 women (mean ± SD age = 22.7 ± 2.1 years; height = 161.1 ± 6.0 cm; body weight = 63.0 ± 11.6 kg) volunteered to participate in this investigation. None of the subjects were competitive athletes, but all were at least recreationally active. The study was approved by the University Institutional Review Board for Human Subjects, and all the subjects completed a preexercise health status questionnaire and signed an informed consent form before testing.
Before the isokinetic testing, the subjects were required to report to the laboratory to be familiarized with the isokinetic dynamometer (LIDO Multi-Joint II, Loredan Biomedical, West Sacramento, CA, USA). During this familiarization session, the subjects were strapped to the dynamometer in accordance with the LIDO user's manual and performed concentric isokinetic muscle actions of the dominant (based on kicking preference) leg extensors at a velocity of 180°·s−1. The first 10 muscle actions were used only as a warm-up, and the subjects were instructed to provide an effort corresponding to approximately 50% of their maximum. After these 10 muscle actions, the subjects were informed regarding the procedures for the fatigue test. Specifically, the subjects were instructed to perform 50 consecutive maximal muscle actions of the leg extensors, and the investigator gave strong verbal encouragement in an attempt to ensure that the subjects were providing a maximal effort during each muscle action. The purpose of the familiarization session was only to familiarize the subjects with the isokinetic dynamometer and fatigue test and to remove any potential influence of a learning effect on isokinetic peak torque. Thus, no data were collected during this session.
At least 48 hours after the familiarization session, the subjects returned to the laboratory to perform the fatigue test in the same manner as that during the familiarization session. All the subjects were instructed to maintain their normal sleeping pattern the night before the testing session, and consistent instructions were provided to each subject in an effort to control arousal levels as much as possible during the testing. After securing the subject into the dynamometer and performing a warm-up set of 10 submaximal muscle actions at 50% of their perceived maximum, the subjects performed the fatigue test. Like the familiarization session, the fatigue test required the subjects to perform 50 consecutive maximal concentric isokinetic muscle actions of the dominant leg extensors at a velocity of 180°·s−1. The subjects were provided with strong verbal encouragement to provide a maximal effort during each muscle action as the data were being collected by the investigator.
The torque signal from the isokinetic dynamometer was sampled continuously throughout the fatigue test by a 12-bit analog-to-digital converter (NI 9201 Voltage Input Module, National Instruments, Austin, TX, USA) at a rate of 3,000 samples per second. The isokinetic peak torque value for each contraction was determined based on the highest value in the torque curve. The initial peak torque was defined as the average from the 3 contractions with the highest peak torque values, and the final peak torque was determined as the average from the 3 contractions with the lowest peak torque values.
During each muscle action of the fatigue test, a bipolar surface EMG signal was detected from the vastus lateralis of the dominant thigh. After carefully shaving and cleansing the skin with alcohol, the electrodes (circular, 1-cm diameter, Ag/AgCl, Ambu, Ballerup, Denmark) were placed over the muscle in accordance with the recommendations from the Surface Electromyography for Non-Invasive Assessment of Muscles (SENIAM) project (8) and approximately in line with the pennation angle of the muscle fibers. The interelectrode distance was 30 mm, and the reference electrode was placed over the patella. After detection, the EMG signals were preamplified (gain = 1,000, Biovision EMG amplifiers, Wehrheim, Germany) and analog filtered (bandpass filter with cutoff frequencies of 10 and 1,500 Hz). The EMG signals were then digitized at a sampling rate of 3,000 samples per second with a 16-bit analog-to-digital converter (Biovision, Wehrheim, Germany) and stored in a personal computer (Dell Optiplex 745, Round Rock, TX, USA) for subsequent analyses.
The EMG signal for each muscle action was first selected from the middle 30° of the range of motion to avoid the acceleration and deceleration phases of movement, which are common with isokinetic dynamometers. These selected signals were then processed with the wavelet analysis described by von Tscharner (15), which uses a filter bank of 11 nonlinearly scaled Cauchy wavelets designed specifically for EMG signals. These Cauchy wavelets are designed to extract the intensities that make up EMG signals and resolve them in both time and frequency. Each wavelet is defined in frequency space by equation 1 shown in Table 1. In this equation,
represents the Fourier transform of the wavelet, f is frequency, f_c is the center frequency of the wavelet, and scale is the scaling factor. The center frequency for each wavelet was calculated using equation 2 shown in Table 1. In this equation, f_cj is the center frequency of wavelet j, and q and r were used to optimize the spacing of the wavelets in frequency space. Specifically, q and r were determined by a least squares approximation such that the sum of all wavelets would be as close as possible to being constant from 20 to 200 Hz. The real and imaginary wavelets have the same power spectrum, and the scale determines the time resolution and bandwidth of the wavelet. The wavelets were indexed by j = 0, 1, 2, …, J, and the total number of wavelets (J + 1) determines the frequency range covered in the wavelet analysis. As described by von Tscharner (15), an optimal set of Cauchy wavelets for EMG can be obtained with J = 10, q = 1.45, and r = 1.959.
Intensity Calculations and Power Wavelets
The Cauchy wavelets developed by von Tscharner (15) are not orthogonal. Thus, the crossproducts of the wavelets have to be considered in the intensity calculation to guarantee energy preservation. This was achieved by multiplying the wavelets with an intensity factor (ck,j), which is defined by equation 3 shown in Table 1. In this equation, k is a frequency index that runs from 0 Hz to one-half of the sampling rate. Multiplication of the Cauchy wavelets with ck,j yields the power wavelets that are used to analyze the signal.
The wavelet analysis was performed in frequency space by multiplying each of the wavelets in frequency space with the Fourier transform of the input signal. The real (Re(wj,k)) and imaginary (Im(wj,k)) wavelet transformed signals were then converted into the time domain with the inverse Fourier transform. The (Re(wj,k)) and (Im(wj,k)) represent the real and imaginary wavelet transformed signals, respectively, for wavelet j throughout the frequency range k. As in the Fourier transform, the intensity (pj,n) (where n is the time index) extracted by each wavelet is the sum of the squares of the real and imaginary parts of the input signal (i.e., [Re(wj,k)]2 + [Im(wj,k)]2). The sum of the intensities for all wavelets provides a measure of the total intensity (Pn) of the signal over time. Total intensity is quantitatively related to traditional measures of power, because Pn is approximately equal to the mean square value of the signal (i.e., the square of the root mean square value) in a given time window.
Time Resolution and Bandwidth
Bandwidth was defined in frequency space as the width of each wavelet at 1/e of its peak intensity (14,15). Time resolution was calculated by first performing the wavelet analysis on the time space representation of the wavelet. Time resolution was defined as the width of the total intensity curve at 1/e of its maximum. Bandwidth, time resolution, and the product of bandwidth and time resolution for each wavelet are shown in Table 2.
Filtering Oscillations Shorter Than the Time Resolution
Oscillations shorter than the wavelet's time resolution often occur in intensity (pj,n) if the rise time of the input signal is shorter than the time resolution. Although these oscillations are partially removed by summing the intensity across all wavelets (i.e., calculating instantaneous intensity), they can be problematic when examining the intensity derived from a single wavelet. Thus, these oscillations do not provide meaningful information about the input signal and were attenuated by a Gauss filter with a width equal to three-eighths of the wavelet's time resolution (15).
Intensity Patterns and Pattern Space
The result of the wavelet analysis is an intensity pattern with time on the x-axis, wavelet index (which determines the frequency) on the y-axis, and intensity on the z-axis (Figures 1–3). Intensity is analogous to power and reflects the fluctuations in the energy of the signal in both time and frequency. The intensity patterns were then stored for subsequent analyses in pattern space. The reader is referred to the original publications (16,19) for details regarding the theory underlying pattern space. We will, however, provide the details of the pattern space calculations that were used in our study. The research question under consideration was whether or not the EMG intensity patterns changed during the fatigue test in a manner that could be predicted and quantified with the trend plot. Thus, a separate pattern space was constructed for each subject by using the EMG intensity patterns from the fatigue test. Each intensity pattern was first consolidated into one long column vector by stacking the intensity values from each wavelet index one on top of the next. This resulted in a matrix (hereafter referred to as the DATA matrix) of 11,000 rows (11 wavelet indices × 1,000 points per index) and 50 columns (50 contractions). The DATA matrix then served as the input for the principal components analysis (aka eigenvector decomposition) (16,19) that was performed without subtraction of the mean. The principal components, or eigenvectors of the covariance matrix, form the axes that span pattern space. The EMG intensity patterns were then projected into pattern space with equation 4 shown in Table 1. In this equation, the PC_axes are the axes that span pattern space, and p_vectors are the locations of the intensity patterns in pattern space. The maximum number of possible dimensions in the principal components analysis is equal to the number of EMG intensity patterns, which was 50 in this study. However, we found that changing the dimensionality had a large influence on the linearity of the resulting trend plot. Specifically, in many cases, using all 50 dimensions added noise to the data and reduced the trend plot linearity. This issue was discussed in detail by von Tscharner (17) and necessitated a decision regarding how many dimensions to use for the principal components analysis. For each subject, we decided to use the fewest dimensions possible that still provided the largest linear correlation coefficient in the trend plot. Thus, the number of dimensions varied across subjects but was always >35.
As discussed by von Tscharner (16), pattern space is difficult to visualize because of its multidimensionality. Thus, trend plots are helpful not only for visualizing movement of the intensity patterns in pattern space but also for analyzing this movement statistically. In this study and that of von Tscharner (16), the trend plot was created by first tracking the movement of the EMG intensity patterns along a path in pattern space. The slope and intercept for this path can then be calculated with linear regression. The slope represents the parts of the EMG intensity patterns that change with time, and the intercept is the expected EMG intensity pattern before the fatigue test. Relative p_vectors are then calculated by subtracting the p_vectors from the intercept and then dividing the result by the magnitude of the intercept. After normalizing the slope with respect to its magnitude, the relative p_vectors are projected onto it, which reflects the percentage change in the magnitude of the p_vectors along the direction of the slope with respect to the magnitude of the intercept (16). These projections can then be plotted on an x-y graph, and a linear regression can be used again to determine if the slope of these projections is significantly different from zero. An alpha level of 0.05 was used to determine statistical significance for the slopes of the trend plots. Previous data from our laboratory have shown high test-retest reliability statistics for concentric isokinetic peak torque (intraclass correlation coefficients ≥ 0.95 with no mean differences between test and retest values).
Table 3 shows the initial peak torque, final peak torque, and percent decline in peak torque for each subject during the fatigue test. The mean ± SD percent decline in the peak torque was 51.6 ± 7.7%. When separated based on gender, the mean ± SD percent decline in peak torque for the men was 50.3 ± 8.3%, whereas that for the women was 53.7 ± 6.8%. Table 4 demonstrates the results from the linear regressions on the trend plots for each subject. All 18 linear regressions demonstrated a slope coefficient that was significantly different from zero (12 were greater, 6 were smaller). The statistical significance of these slope coefficients was not related to the percent decline in peak torque (i.e., a large percent decline in peak torque did not guarantee a statistically significant linear slope coefficient). There were also no gender differences in the linear slope coefficients for the trend plot. Figure 1 shows the EMG intensity patterns for the first and last contractions, and the trend plot for subject 15 during the fatigue test. Figures 2 and 3 demonstrate the same EMG intensity patterns and trend plots for subjects 1 and 18, respectively. Note that in each of these figures, the changes that occurred in the EMG intensity pattern from the first contraction to the last contraction could not be simply described by a shift of intensity to lower frequencies. Instead, the overall structure of the patterns changed, and this finding was consistent for all the subjects. Figures 1–3 also demonstrate the between-subject differences in the overall structure of the EMG intensity patterns. There are no 2 patterns that are alike when comparing different subjects, which illustrates the uniqueness and specificity of EMG intensity patterns.
The results from this study indicated that in all the cases (18 out of 18 subjects), the EMG intensity patterns moved in a predictable manner in pattern space. Specifically, their path of movement could be tracked with the trend plot and predicted with the resulting linear regression line. The finding that this movement was statistically significant for all the subjects is most likely a reflection of the severity of the fatigue experienced by the subjects. As shown in Table 3, the average percent decline in isokinetic peak torque during the fatigue test was just >50%. Despite this large decrease in peak torque, the changes that occurred in the EMG intensity pattern were not consistent across the subjects and could not be characterized simply by a shift of intensity to lower frequencies.
Von Tscharner (16) first developed the concepts of pattern space and the trend plot for the purpose of investigating fatigue during very low-intensity cycling. The author (16) reported that in many cases (up to 80% or more, depending on the muscle examined), the slope of the trend plot was statistically significant. This finding was important because it demonstrated the sensitivity of the EMG signal and the trend plot to fatigue that might not have even been perceptible by the subjects. Our findings also indirectly supported those from a recent study by Stirling et al. (13), who reported differences in the EMG intensity patterns between the low- and high-effort stages of prolonged running. Specifically, EMG signals were detected from several muscles of the thigh and lower leg during a 1 hour treadmill run in which the subjects provided ratings of perceived exertion every 6 minutes to quantify their effort level. The EMG signals were then processed with the same wavelet procedure used in this study and the resulting intensity patterns were classified using a support vector machine approach. The authors (13) found that the patterns could be classified into their respective low- and high-effort level groups with accuracy rates that were always >80%. Thus, it was concluded that when combined with the wavelet analysis and pattern classification techniques, EMG could potentially be used as a sensitive tool for monitoring the fatigue status of the muscles while athletes are training (13). The findings from this study also supported this hypothesis, and indicated that although the changes to EMG intensity patterns are complex, they do occur in a predictable manner that can be monitored with the trend plot.
It would be tempting to assume that the changes to the EMG intensity patterns in this study were simply a function of a shift in intensity to lower frequencies. Lindström et al. (10) discussed the fact that these low frequency shifts were mostly because of decreases in conduction velocity with fatigue. In fact, a preliminary analysis of our data indicated a consistent decrease in EMG mean frequency during the fatigue test for all subjects. It is important to remember, however, that the trend plot is capable of resolving even very minor changes in EMG intensity patterns. Thus, any change to the overall structure of the EMG intensity pattern, whether it be a change in total intensity, timing, or frequency can be resolved with the trend plot. It is also very possible that the changes to the EMG intensity patterns in this study reflected rhythmic bursts of synchronized motor unit activity, often referred to as the “Piper Rhythm” (18). von Tscharner et al. (18) recently used wavelet-based techniques to identify pacing strategies in EMG signals from the abductor pollicis brevis muscle during repeated maximal isometric muscle actions. The authors (18) found rhythmic bursts in the EMG signals as the muscle became fatigued that were hypothesized to reflect synchronous motor unit activity. These rhythmic bursts were first noticed by Piper (11) and may reflect a specific control strategy by the central nervous system in response to fatiguing activity. Although the occurrence of the Piper Rhythm has not been examined during dynamic muscle actions, it is possible that it affected the changes observed in the EMG intensity patterns in this study. With advances in signal processing, future investigations may be able to determine if these synchronous bursts occur during dynamic muscle actions, and, if so, what their magnitude and frequency is.
Even though this study showed consistent within-subject changes in the EMG intensity patterns, the differences between subjects were not consistent. As shown in Figures 1–3, the changes to the overall structure of the EMG intensity pattern with fatigue were very complex and likely reflective of between-subject differences in factors such as fiber type composition, pennation angle, and possibly even muscle size. These morphological differences all have the potential to influence the EMG intensity pattern in a subject-specific manner. This does not, however, limit the potential for using EMG intensity patterns and the trend plot to examine muscle fatigue. The results from this study and those of von Tscharner (16) demonstrated that the within-subject changes to EMG intensity patterns were consistent and could be described with the trend plot. The fact that there were large between-subject differences, however, means that definitive conclusions cannot be made regarding the effects of fatigue on the EMG signal during dynamic muscle actions.
In summary, our findings indicated that when combined with wavelet analysis of EMG signals, the trend plot is a sensitive indicator of muscle fatigue. All the 18 subjects demonstrated linear slope coefficients for the trend plot that were significantly different from zero. There were, however, large between-subject differences in the changes to the EMG intensity patterns. Thus, definitive conclusions regarding the effects of fatigue on EMG intensity patterns cannot be made. The trend plot does, however, combine the intensity, timing, and frequency aspects of EMG signals into a single parameter. Thus, it is very rich with information and can potentially be used to monitor the fatigue status of a muscle or muscle group. Future studies need to be carried out to determine the sensitivity of the trend plot for monitoring fatigue in different muscles and during various types of activities.
This study showed that the fatigue that develops during repeated dynamic muscle actions can be tracked with wavelet analysis of EMG signals and a trend plot. These tools are capable of analyzing subtle changes in EMG signals and are much more sensitive than simple amplitude or center frequency parameters. Thus, these techniques could be very useful for coaches or athletes interested in tracking the fatigue status of individual muscles. Specifically, consistent movement of the patterns within the trend plot reflects the systematic changes in the amplitude, frequency, and timing of the events that make up the EMG signal. These events provide detailed information regarding the muscle's activation pattern, and, therefore, show the unique strategies used by each subject for coping with fatigue. Potential applications of these methods include tracking fatigue during any endurance-based activity, such as cycling, running, or rowing. Strength and conditioning professionals might also be able to use the methods to examine fatigue in individual muscles during repetitive multijoint activities, such as squats and power cleans.
The authors declare no personal or financial relationships with companies or manufacturers who may benefit from the results of this study. The results from this study do not constitute endorsement from the National Strength and Conditioning Association.
1. Basmajian JV, De Luca CJ. Muscles Alive: Their Functions Revealed by Electromyography (5th ed.). Baltimore, MD: Lippincott Williams and Wilkins, 1985.
2. Bonato P, Gagliati G, Knaflitz M. Analysis of myoelectric signals recorded during dynamic contractions: A time-frequency approach to assessing muscle fatigue
. IEEE Eng Med Biol Mag Nov/Dec: 102–111, 1996.
3. Bonato P, Roy SH, Knaflitz M, De Luca CJ. Time-frequency parameters of the surface myoelectric signal for assessing muscle fatigue
during cyclic dynamic contractions. IEEE Trans Biomed Eng 48: 745–753, 2001.
4. Clancy EA, Bertolina MV, Merletti R, Farina D. Time- and frequency-domain monitoring of the myoelectric signal during a long-duration, cyclic, force-varying, fatiguing hand-grip task. J Electromyogr Kinesiol 18: 789–797, 2008.
5. Cobb S, Forbes A. Electromyographic studies of muscle fatigue
in man. Am J Physiol 65: 234–251, 1923.
6. Englehart K, Hudgins B, Parker PA, Stevenson M. Classification of the myoelectric signal using time-frequency based representations. Med Eng Phys 21: 431–438, 1999.
7. Farina D, Lucas MF, Doncarli C. Optimized wavelets
for blind separation of nonstationary surface myoelectric signals. IEEE Trans Biomed Eng 55: 78–86, 2008.
8. Hermens HJ, Freriks B, Merletti R, Stegeman D, Blok J, Rau G, Disselhorst-Klug C, Hägg G. SENIAM European Recommendations for Surface ElectroMyoGraphy: Results of the SENIAM project. Enschede, The Netherlands: Roessingh Research and Development, 1999.
9. Karlsson JS, Gerdle B, Akay M. Analyzing surface myoelectric signals recorded during isokinetic contractions. IEEE Eng Med Biol Mag 20: 97–105, 2001.
10. Lindström L, Magnusson R, Petersén I. Muscular fatigue
and action potential conduction velocity changes studied with frequency analysis of EMG signals. Electromyography 10: 341–356, 1970.
11. Piper H. Ueber den willkuerlichen Muskeltetanus. Pfluegers Gesamte Physiol Menschen Tiere 119: 301–338, 1907.
12. Rioul O, Vetterli M. Wavelets
and signal processing. IEEE Sig Proc Mag Oct: 14–38, 1991.
13. Stirling LM, von Tscharner V, Kugler PF, Nigg BM. Classification of muscle activity based on effort level during constant pace running. J Electromyogr Kinesiol 21: 566–571, 2011.
14. Torrence C, Compo GP. A practical guide to wavelet analysis. Bull Am Meteor Soc 79: 61–78, 1998.
15. von Tscharner V. Intensity analysis in time-frequency space of surface myoelectric signals by wavelets
of specified resolution. J Electromyogr Kinesiol 10: 433–445, 2000.
16. von Tscharner V. Time-frequency and principal-component methods for the analysis of EMGs recorded during a mildly fatiguing exercise on a cycle ergometer. J Electromyogr Kinesiol 12: 479–492, 2002.
17. von Tscharner V. Spherical classification of wavelet transformed EMG intensity patterns. J Electromyogr Kinesiol 19: 334–344, 2009.
18. von Tscharner V, Barandun M, Stirling LM. Fatigue
-related decrease in piper rhythm frequency of the abductor pollicis brevis muscle during isometric contractions. J Electromyogr Kinesiol 21: 190–195, 2011.
19. von Tscharner V, Goepfert B. Gender dependent EMGs of runners resolved by time/frequency and principal pattern analysis. J Electromyogr Kinesiol 13: 253–272, 2003.
20. von Tscharner V, Valderrabano V. Classification of multi muscle activation patterns of osteoarthritis patients during level walking. J Electromyogr Kinesiol 20: 676–683, 2010.
Keywords:© 2012 National Strength and Conditioning Association
neuromuscular; wavelets; fatigue