Automatic identification of motor unit action potential trains from electromyographic signals using fuzzy techniques E. Chauvet1
O. Fokapu1,2
J.-Y. Hogrel3
D. Gamet1
J. Ducheˆne4
1
LBIM, UMR CNRS 6600, Universite´ de Technologie de Compie`gne, Compie`gne, France 2 Universite´ de Picardie Jules-Verne, Amiens, France 3 Institut de Myologie, GH Pitie´-Salpeˆtrie`re, Paris, France 4 LM2S, Universite´ de Technologie de Troyes, Troyes, France
Abstract—A technique is proposed that allows automatic decomposition of electromyographic (EMG) signals into their constituent motor unit action potential trains (MUAPTs). A specific iterative algorithm with a classification method using fuzzy-logic techniques was developed. The proposed classification method takes into account imprecise information, such as waveform instability and irregular firing patterns, that is often encountered in EMG signals. Classification features were determined by the combining of time position and waveform information. Statistical analysis of inter-pulse intervals and spike amplitude provided an accurate estimation of features used in the classification step. Algorithm performance was evaluated using simulated EMG signals composed of up to six different discharging motor units corrupted with white noise. The algorithm was then applied to real signals recorded by a high spatial resolution surface EMG device based on a Laplacian spatial filter. On six groups of 20 simulated signals, the decomposition algorithm performed with a maximum and an average mean error rate of 2.13% and 1.37%, respectively. On real surface EMG signals recorded at different force levels (from 10% to 40% of the maximum voluntary contraction), the algorithm correctly identified 21 MUAPTs, compared with the 29 MUAPTs identified by an experienced neurophysiologist. The efficiency of the decomposition on surface EMG signals makes this method very attractive for non-invasive investigation of physiological muscle properties. However, it can also be used to decompose intramuscularly recorded EMG signals. Keywords—Decomposition, Fuzzy logic, Motor unit, Electromyographic signal Med. Biol. Eng. Comput., 2003, 41, 646–653
1 Introduction WAVEFORMS AND firing rates of motor unit action potentials (MUAPs) in an electromyographic (EMG) signal provide an important source of information for physiological and clinical investigations of the neuromuscular system (BASMAJIAN and DE LUCA, 1985; IANI et al., 1994). Such an investigation is possible if identified MUAP trains (MUAPTs) are available. Several decomposition techniques can be found throughout the literature (LEFEVER and DE LUCA 1982a;b; MCGILL et al., 1985; CHRISTODOULOU and PATTICHIS, 1999; FANG et al., 1999). However, these various approaches process mainly intramuscular EMG signals that are detected using invasive electrodes. The identification of single MUAPs from surface EMG recordings has been shown to be possible (ROELEVELD et al., 1997), especially from high-resolution EMG systems (RAU and Correspondence should be addressed to Mr Eric Chauvet; email:
[email protected] Paper received 29 November 2002 and in final form 30 June 2003
MBEC online number: 20033817 # IFMBE: 2003
646
DISSELHORST-KLUG, 1997). However, the proposed methods do not allow the evaluation of MU firing patterns. To our knowledge, the decomposition of surface EMG signals into their constituent MUAPTs has yet to be successfully demonstrated. A recording set-up based on a spatial-filtering multi-electrode array reduces interferences in the surface EMG signal, making MUAPT identification feasible according to an EMG decomposition concept (REUCHER et al., 1987a;b; HOGREL and DUCHEˆNE, 1999). Therefore the aim of our work was to develop an automatic decomposition method for the investigation of single MUs, with specific focus on the applicability to surface EMG signals. Many processes that decompose an interferential, intramuscular EMG signal into its constituent MUAPTs have already been developed. A review of these techniques can be found in STASHUK (2001). In general, basic processing of any intramuscular decomposition method involves three sequential stages: (i)
segmentation of the composite signal: the EMG signal is cut into segments of isolated spike potentials (ii) feature extraction: for pattern recognition, where the detected spikes are represented using a set of relevant characteristics Medical & Biological Engineering & Computing 2003, Vol. 41
(iii) classification of detected spikes: the detected MUAPs are distributed into different clusters based on MUAP similarities. Nevertheless, the inherent differences between intramuscular and surface EMG signals, mainly due to the interferential nature of surface recordings, are such that methods developed for intramuscular EMG are not necessarily well-adapted for surface EMG signals. For these reasons, we considered it relevant to develop a decomposition method (a) taking into account the specific characteristics of the surface EMG signal (b) robust enough to account for signal variability (c) requiring a low computational cost (d) applicable to both intramuscular and surface EMG signals. In this work, a novel decomposition approach for EMG MUAPT identification is introduced. A specific iterative decomposition algorithm, with a classification method based on fuzzy logic, is proposed. The aim of the iterative principle is to reduce the number of MUAPs simultaneously under identification. A fuzzy-based classification uses input variables derived from firing pattern and waveform information to refine the identification of each MUAPT. The presented algorithm is able to decompose an EMG signal at low to moderate force levels, where the number of detectable MUAPTs is less than six.
2 Methods The decomposition algorithm based on an iterative principle includes three main steps in each iteration: signal segmentation; feature extraction based on statistical analysis; and classification using fuzzy logic. Each iteration can lead to the identification of, at most, one MUAPT. These three steps were included in a global loop based on successive amplitude thresholds estimated at the beginning of each iteration. The flowchart of the overall decomposition algorithm is given in Fig. 1. 2.1 Signal segmentation At a given iteration, spikes were detected when their amplitudes were higher than a detection threshold value. At the first iteration, the detection threshold was initialised at the maximum amplitude of the signal segment under study. After thresholding, the number of detected spikes was counted. If this number did not reach at least 5 spikes per second, according to the lower estimated MU firing rates (BASMAJIAN and DE LUCA, 1985), the threshold level was lowered to 90% of its previous value. To isolate the detected spikes, a 10 ms window, centred on the maximum of each detected spike, was applied. A shorter length can miss the main part of the MUAP spike in the case of surface EMG signals, as MUAPs are usually of longer duration than intramuscular EMG MUAPs (BASMAJIAN and DE LUCA, 1985). All stored spike segments were aligned with respect to their peak position. The occurrence time of the peak of each spike was also stored. The principle is illustrated in Fig. 2. 2.2 Feature extraction For MUAPT identification, three parameters were used to build the feature space: inter-pulse interval (IPI), peak-to-peak amplitude and the MUAP waveform. From all detected spikes, the algorithm searched for a representative IPI IPIrep and a representative peak-to-peak amplitude PPArep; these were subsequently used to select a representative MUAP waveform Mwrep. Medical & Biological Engineering & Computing 2003, Vol. 41
Fig. 1 Flowchart of surface EMG decomposition algorithm
The estimation of a relevant value of IPIrep and PPArep was based on the distribution function of the inter-pulse intervals and peakto-peak amplitudes, respectively. 2.2.1 Search for a representative IPI IPIrep: The procedure consisted of two phases: Initialisation: Inter-pulse intervals between all pairs of detected spikes were computed. The IPI distribution function was built and restricted to the range of IPIs deemed to be physiologically relevant (BASMAJIAN and DE LUCA, 1985). The cumulative distribution function of the remaining IPIs was calculated. Estimation: After re-sampling (16104 Hz), the first derivative of the distribution function was computed so that we could estimate the probability density function. To reduce fluctuations, a zerophase filter was applied on the derivative 1 [(x xn3 ) þ 4(xnþ2 xn2 ) 32 nþ3 þ 5(xnþ1 xn1 )]
y(n) ¼
(1) 647
Initialisation: The peak-to-peak amplitudes of the spike potentials of the valid IPIs were computed, and the corresponding distribution function was constructed. Estimation: As for IPIrep, PPArep was deduced from the smoothed derivative of the peak-to-peak amplitude distribution function. An illustration of the estimation of IPIrep and PPArep is provided in Fig. 3. 2.2.3 Search for representative MUAP waveform Mwrep: The aim was to determine the spike that best represented the MUAP waveform from the set of selected spike potentials. For this purpose, two steps were necessary: Initialisation: The whole set of spikes for the valid IPIs was first considered. Estimation: The spike with the peak-to-peak amplitude closest to PPArep was selected as the representative MUAP waveform. IPIrep, PPArep (scalar values) and Mwrep (1D array) constituted the reference set for further classification.
Fig. 2
Detection and segmentation of real surface EMG signal. (a) Original signal and threshold level at end of first iteration. (b) Spike segments after alignment
where xn represents the interpolated data of the IPI distribution function. The IPIrep value was calculated as the value of xn that produced the maximum value of y(n). An additional selection of the IPIs to be kept was performed by the removal of IPI values that differed from IPIrep by more than 20% and were considered as irregular firing intervals (LEFEVER and DE LUCA, 1982a; MCGILL et al., 1985; STASHUK and QU, 1996b). 2.2.2 Search for a representative peak-to-peak amplitude PPArep: The procedure was similar to IPIrep estimation. Peak-to-peak amplitude was defined as the difference between the positive and negative peaks. 648
Fig. 3 Determination of representative (a) IPI and (b) PPA. (- - -) Interpolated distribution function. (—) Smoothed derivative. Principle is applied on spike potentials shown in Fig. 2 Medical & Biological Engineering & Computing 2003, Vol. 41
2.3 Fuzzy logic-based classification This step was intended to select the potentials belonging to the same MUAPT. Identification of MUAPTs from surface EMG signals can be achieved using a fuzzy approach for MUAP classification purposes. ZADEH (1965) introduced the theory of fuzzy sets, where the membership of objects to classes is a matter of degree. The conventional structure of a fuzzy-logic system (FLS) is composed of four components, namely the fuzzifier, the fuzzy rule base, the inference engine and the defuzzifier. Details of FLS implementation procedures can be found in MAMDANI and ASSILIAN (1975), ZADEH (1978) and TAKAGI and SUGENO (1985). The FLS procedure started with the set of detected spikes and the reference parameters obtained previously. The classification process had an iterative structure so that no more than one MUAPT was identified per iteration.
Table 1 Decision table representing rules. Small (S), moderate (M) and large (L) are linguistic values of input fuzzy variables. A (accept current MUAP) and R (reject current MUAP) are fuzzy partial consequences. Each rule can be linguistically written as follows: IF DIPI is Large AND DPPA is Large AND C is Small THEN Reject the current MUAP. IF DIPI is Small AND DPPA is Moderate AND C is Large THEN Accept the current MUAP DIPI
DPPA
L M S
Input=output fuzzy-variable definition: derived fuzzy-logic parameters: Three variables were defined as inputs for the fuzzy classifier. The first input was the linguistic variables ‘relative deviation in terms of IPI’ DIPI
C
S
M
L
S M L S M L S M L
R R A R A A A A A
R R R R A A R A A
R R R R R R R R R
(2)
and the correlation C is higher. The satisfaction degree of the fuzzy partial consequence (A ¼ accept the current MUAP; R ¼ reject the current MUAP) was computed for each rule.
where IPIcur was the inter-pulse interval between the current MUAP to be classified and the last accepted MUAP. DIPI represents the membership degree of the current MUAP to the class according to its time position in the signal. The second input was the linguistic variables ‘relative deviation in terms of amplitude’ DPPA
The inference engine: A ‘symbolic fuzzy inference engine’ was used. The conjunction was modelled by the ‘Min’ operator and, for the modelling of fuzzy partial consequences aggregation, the ‘Max-Union’ operator was used.
DIPI ¼
jIPIcur IPIrep j IPIrep
DPPA ¼
jPPAcur PPArep j PPArep
(3)
where PPAcur was the peak-to-peak amplitude of the current MUAP to be classified. DPPA represents the membership degree of the current MUAP to the class according to its peak-to-peak amplitude. The third input was the linguistic variable ‘correlation’ C. It represents the membership degree of the current MUAP according to its shape, i.e. the value of the correlation coefficient between the current MUAP and the MUAP selected as Mwrep. The outputs of the system consisted of two linguistic variables. The first was a satisfaction degree for the current MUAP to belong to the class, SDB. The second was a satisfaction degree for the current MUAP to be rejected, SDR. Fuzzy partition qualifying the input variables: The border values of the trapezoidal fuzzy intervals are represented in Fig. 4. They were determined according to the literature (BASMAJIAN and DE LUCA, 1985; IANI et al., 1994). Setting up fuzzy rules: The decision table defines a system of 27 fuzzy rules (Table 1). The strategy to build the decision table relied on the following principle: the current MUAP is more acceptable when the relative deviations (DIPI and DPPA) are lower
Defuzzification: The partial conclusions were represented by the value of the two variables SDB and SDR. The final conclusion was determined by taking into account the two preceding values and the following principle: the higher the SDB and the lower the SDR, the more acceptable the current MUAP. 2.4 Residual signal for the next iteration Before the next iteration for a new MUAPT identification, the MUAPT identified in the preceding step had to be removed from the EMG segment. Subtraction of the reference MUAP shape often created artifacts on the remaining EMG signal. These artifacts seriously reduced the performance of subsequent iterations. To overcome this problem, the corresponding MUAP segments were simply replaced by zeros. This led to a reduced signal called the ‘residual signal’, which was used for the next iteration. The detection threshold was lowered at each iteration. The loop stopped when the threshold reached the segment RMS value, which guaranteed that segments containing only noise were not considered. 2.5 Merging of MUAPTs This step was needed when MUAPs belonging to the same MU had been classified as two different MUAPTs. The candidates for merging had to show similar MUAP shapes and
Fig. 4 Membership functions of three input variables. Small (S), Moderate (M) and Large (L) are linguistic values of input fuzzy variables Medical & Biological Engineering & Computing 2003, Vol. 41
649
compatible firing rates. In that way, two conditions were necessary to decide the merging: (i)
high correlation: the correlation coefficients between all pairs of MUAP waveforms were computed. The correlation coefficient had to be higher than 0.7. (ii) consistency of the new firing-rate: the algorithm searched for the IPIrep after merging using the principle described in Section 2.2.1. The assumption was made that at least 95% of new IPI values must be located inside the IPI range (IPIrep 20%).
If these two conditions were simultaneously respected, the two MUAPTs were merged. A new Mwrep was then obtained by averaging the two Mwreps of the initial classification. 2.6 Reference signals for evaluation of the decomposition methods Simulated and real surface EMG signals were used to evaluate the algorithm. The main idea behind this evaluation was to assess the capability of the algorithm to separate an interferential signal into as many MUAPTs as possible. 2.6.1 Simulated signals: These were produced by the simulation software developed by DUCHEˆNE and HOGREL (2000). This software is based on a model taking into account the main parameters influencing EMG signal characteristics. This model includes all transformations from intracellular potential to surface recordings. Each signal was simulated by superimposition of between one and six individual MUAPTs. IPIs were randomised by a Gaussian distribution (m ¼ 90 ms, s ¼ 25 ms, range ¼ 40–140 ms). The IPI variation was extracted from a zero-mean Gaussian distribution, the standard deviation of which was computed as 10% of the mean IPI. Each segment of the simulated EMG recordings was 3 s long and was corrupted by Gaussian white noise (SNR ¼ 20 dB). Six groups of 20 segments each were generated, the first group with only one MUAPT, the second with two, and so on up to six. 2.6.2 Real EMG signals: Surface EMG signals were recorded using a Laplacian system developed by HOGREL and DUCHEˆNE (1999). This recording configuration acts as a spatial filter on the potential distribution at the skin surface. The Laplacian configuration is composed of five electrodes. The central electrode is weighted by a factor of 4, and the four surrounding electrodes are weighted by a factor of þ1. The electrode set was composed of 11 electrodes producing three spatially filtered EMG channels (Fig. 5). Signals were filtered and pre-amplified within the electrode unit and then sent to an amplifier (frequency band: 10–1000 Hz). Then the signals were sampled at 10 kHz with a 12-bit A=D converter. Two healthy subjects performed isometric contractions of the biceps brachii at various levels, corresponding to approximately 10%, 20%, 30% and 40% of their maximum voluntary contraction (MVC). A set of 5 s tests were performed at each contraction level. A total of 16 recordings (four per level) were visually analysed to extract 3 s segments corresponding to their stationary part.
2.7 Evaluation criteria The interest of an evaluation by simulated signals was that the individual MUAPTs constituting the signal were known a priori and could be compared with the result of the decomposition algorithm. 650
Fig. 5 High-resolution system for surface EMG detection. (a) Laplacian configuration of electrodes. (b) Three-channel Laplacian acquisition device. (c) Example of signals recorded by three channels
The efficiency of the method was evaluated by the following four criteria: (a) the IPI accuracy rate, which referred to the percentage of relative deviation between simulated IPI value and the value obtained using the identified MUAPT (b) the MUAP assignment error, which referred to the percentage of the total number of MUAPs that were wrongly assigned to a train (c) the extent, which referred to the detection rate of MUAPs within a train (d) the MUAPT detection rate, which referred to the total number of identified MUAPTs with respect to the total number of simulated MUAPTs. For evaluation from real surface EMG signals, the results produced by the algorithm were compared with MUAPTs identified by an experienced neurophysiologist.
3 Results 3.1 Algorithm results on simulated signals The results are summarised in Table 2. The result was obviously perfect for the first group of signals (100% IPI accuracy, 100% MUAPs assigned, 100% extent, 0% error rate and 100% identified train rate). The IPI accuracy rate value was at least 98.8% for all groups of signals. On average, only 55.6% of the occurring MUAPs were assigned. This was mainly owing to two factors: first, the presence of small MUAPs of similar waveform; and secondly, the presence of superimposed MUAPs that could have been detected as single MUAPs, particularly for the group of six MUs. The average detection rate within a train (extent) was 94.5%. As the interference level increased, the MUAPT detection rate decreased significantly, whereas the MUAP assignment error rate remained low. The decreasing MUAPT detection rate may be related to the increasing superimposition of spikes in the signal. 3.2 Algorithm results on real surface EMG signals Surface EMG signals were detected by a high spatial resolution EMG device at moderate contraction levels. In such Medical & Biological Engineering & Computing 2003, Vol. 41
Table 2 Performances on simulated surface EMG signals Simulated data
Number of MUs 1 2 3 4 5 6
Mean number of MUAPs (range) 29.4 (24–37) 55.9 (49–74) 103.1 (96–110) 143.4 (138–151) 194.8 (158–215) 244.9 (229–254)
Decomposition results
Mean IPI, ms (range)
Mean number of classified MUAPs (range)
Mean accuracy of IPI, % (range)
Mean MUAPT detection rate, % (range)
Mean MUAP assignment error rate, % (range)
Mean extent, % (range)
98.7 (67.7–127.5) 82.3 (65.1–104.8) 87.7 (65.1–106.1) 86.7 (65.7–119.1) 96.2 (82.8–119.1) 94.1 (82.8–119.1)
29.4 (24–37) 52.1 (44–65) 75.3 (46–87) 67.3 (38–78) 64.9 (38–102) 53.9 (35–101)
100 (100–100) 99.6 (98.5–100) 98.8 (86.8–100) 99.6 (99.1–100) 98.8 (80.4–100) 99.6 (98.7–100)
100 (—) 95 (50–100) 78.3 (33.3–100) 45 (25–50) 32 (20–40) 20 (16.7–33.3)
0 (—) 0.86 (0–2.7) 1.34 (0–4.17) 1.85 (0–3.8) 2.03 (0–4.76) 2.13 (0–5.33)
100 (100–100) 97.7 (94.2–100) 87.6 (75.5–96.7) 91.5 (86.6–100) 95.2 (91.3–100) 94.9 (86.9–100)
measurement conditions, with a reduced detection volume, few significant MUAPTs are generally able to be detected at moderate contractions (rarely more than ten MUAPTs). From all recorded segments, 29 MUAPTs were identified by an experienced neurophysiologist. Using the present algorithm, 21 MUAPTs were correctly identified. An example of the iterative decomposition is illustrated in Fig. 6. In this case, for the population of MUAPs assigned at the first iteration, one MUAP with a large amplitude was a superimposed MUAP. The assignment of this MUAP by the fuzzy system was due to its perfect position in the train. Gaps observed in the firing time plot of the last three MUAPTs were due to superimposed waveforms and low spike amplitudes in comparison with noise level. Altered MUAPs were generated by the assigned MUAPs being removed. From the MUAPTs provided by the decomposition and their corresponding IPIrep, an estimation of the mean IPI value was made as follows: (i) computation of the consecutive IPIs (ii) selection of IPI values located inside the range IPIrep 20% (iii) estimation of their mean values. During voluntary isometric contraction, a significant increase in firing rate was observed with the increase in force (from 9.9 1.3 Hz at 10% MVC to 17.0 2.2 Hz at 40% MVC). The relationship between the mean IPI value and the force level was confirmed.
4 Discussion Even with a selective recording system, surface EMG MUAPs are more interferential than intramuscular ones, and superimposed MUAPs are more likely to occur. To take this fact into account, we proposed a specific decomposition procedure that considered the interferential nature of surface EMG signals, while remaining valid for intramuscular signals. The proposed algorithm included the three main steps of a classical decomposition scheme (segmentation, feature extraction, classification). However, to reduce the classification errors, we introduced an iterative procedure focusing on the optimisation of the number of the spikes to be simultaneously processed. Considering the technical aspects of the proposed iterative algorithm, some remarks about the three main steps can be made. Medical & Biological Engineering & Computing 2003, Vol. 41
4.1 Spike detection The purpose was the detection of a set of MUAPs with a high probability that they belonged to the same MUAPT for the current iteration. For threshold setting, there is no standardised procedure for estimating the threshold level. A summary of different approaches can be found in STASHUK (2001). The use of a moving threshold obtained by the lowering of its previous value is an efficient way to improve the detection phase. However, repeating the detection phase until a sufficient number of spikes is obtained can appear time consuming. During our tests, it was observed that the threshold was lowered no more than three times before the feature extraction phase. Even though the detection stage required time, this was compensated for by the time saving obtained during the classification stage.
4.2 Feature extraction Our decomposition algorithm was based on waveform and temporal information. Decompositions that use MUAP shape and MU firing-pattern-based criteria already exist in the literature (LEFEVER and DE LUCA, 1982a;b; STASHUK and QU 1996a; FANG et al., 1999). It has been shown that these algorithms are more efficient than algorithms based on shape criteria only. In this study, the determination of classification features differed from others on two points: First, feature information (firing-pattern and shape information) was obtained from the distribution function of the detected spikes. Firing-rate estimation is usually based on an IPI histogram and thus requires a precise definition of regions containing valid IPIs for an accurate estimation (STASHUK and QU, 1996b). The use of the numerical derivative of the distribution function of detected spikes presented in this study is a simple method to derive the classification parameters. A restrictive point is that the representative IPI may correspond to an inaccurate estimation of mean firing rate, owing to the reduced number of detected spikes in the current iteration, which can lead to duplicated MUAPTs. However, this drawback is compensated for by the merging step. Secondly, the MUAP template is a fusion of firing rate and peak-to-peak amplitude information. In contrast, MUAP templates are classically obtained by the averaging of the MUAPs grouped in an initial class (STASHUK and QU, 1996a; CHRISTODOULOU and PATTICHIS, 1999; FANG et al., 1999; STASHUK, 1999). 651
system avoids hard thresholds, which seems to be an advantage when biological signals are analysed. The characteristics of the recordings can differ widely from subject to subject, as well as for the same subject, owing to different positions, activity levels or fatigue. These characteristics can also be affected by the influence of background noise. In addition, classifiers based on fuzzy logic support the implementation of rules based on the knowledge of a human expert and can be implemented at low computational cost. Classification parameters are those usually used by many decomposition algorithms (LEFEVER and DE LUCA, 1982a;b; STASHUK and QU, 1996a; FANG et al., 1999; STASHUK, 1999). The originality of our method resides in the combination of these parameters. The allocation of a MUAP to a class simultaneously takes into account the three basic criteria (IPI, amplitude and waveform). This combination reduces erroneous classification. Furthermore, the strategy used to build the decision rules leads to a correct assignment of atypical MUAPs (Fig. 6c), as their ?DIPIs are small (IF DIPI is Small AND DPPA is Moderate AND C is Large THEN Accept the current MUAP). Consequently, the number of correctly assigned MUAPs increased. Using this method, gaps in the MUAPT are reduced, which is an advantage for the estimation of firingpattern statistics (mean and standard deviation of the IPIs). Fine tuning of the fuzzy classifier is difficult as the effects of changing membership functions and adding or removing rules are hard to predict when operating on a non-linear system. Most false-negative classifications occurred when the EMG segments to be decomposed contained irregular inter-pulse intervals owing to superimposed MUAPs. Some of the false-positive classifications resulted from only small differences in the shape and time position of two MUAPs. Additional fuzzy input variables would potentially increase the algorithm performance. A simultaneous analysis of multiple channels could also add information, as changes in the shape of the waveform due to other MUAPs may not occur to the same extent in each channel. 4.4 Overall efficiency of the fuzzy logic based method on real signals
Fig. 6
Decomposition of real surface EMG signal, recorded during isometric contraction (20% MVC). (a) 3 s duration of original signal. (b) Representative waveforms. (c) Shimmer plots of assigned MUAPs. (d) Firing time plots of each identified MUAPT. No MUAPT was identified at iteration 2
4.3 Classification Before the decomposition of a real EMG signal, the number of MUAP classes, the number of MUAPs per class, MUAP firing rate and MUAP waveforms are unknown. Moreover, there exists substantial MUAP waveform variability and MUAP superposition in surface EMG signals. Fuzzy-logic techniques are well adapted to the processing of imprecise information, such as surface EMG characteristics. Classification using a fuzzy-logic 652
The proposed decomposition method was applied to a surface EMG signal recorded with moderate interferential level at isometric voluntary contraction. With some adaptation, it could be applied to signals recorded during an isometric contraction ramp. For the exploitation of the decomposition results in term of MUAP parameter measurement (amplitude, duration, area, rise time, phases, turns etc.), a prototypical MUAP of each MU must be determined. This prototypical MUAP could be estimated from the original EMG signal using the spike triggering averaging technique, based on MUAP firing times obtained from identified MUAPTs. The method presented was developed to decompose EMG signals at low to moderate force levels. Among the real signals analysed, the algorithm was able correctly to identify up to four MUAPTs. Concerning signals with high interference levels, as with many decomposition algorithms, the proposed method presented some limitations. Signal decomposition complexity was not only directly related to the level of muscle activity. It was also closely related to the relative time positions and relative amplitudes of the MUAPs within the signal. Hence, firing time and amplitude statistics can enhance signal decomposition complexity, even for signals comprising few MUs. However, it was observed that, during strong voluntary contractions, identification of every MUAPT was not possible. At force levels higher than 40% MVC, the algorithm was not able accurately to identify more than two MUAPTs owing to the superimposition of MUAPs. In this kind of signal, the spikes with small amplitude presented similar waveforms and were Medical & Biological Engineering & Computing 2003, Vol. 41
often obscured by large spikes. Substantial decomposition errors may be due to the wrong identification of small spikes. An effective way to circumvent such a problem is to increase the detection threshold in the signal-detection procedure so that small spikes are considered as noise. Another solution could be to enhance the spatial selectivity of the recording system.
5 Conclusions In this work, an algorithm was proposed to perform correct MUAP classification despite shape variability and background noise, as often encountered in surface EMG signals. Based on this approach, the MUAPT identification procedure provided an alternative technique that enabled MU characterisation even for signals recorded non-invasively. This new decomposition technique has been successfully applied to surface EMG signals recorded from healthy subjects. The algorithm now needs to be tested on signals recorded intramuscularly. The algorithm was designed as a first-level algorithm that did not treat the problem of superimposition. Therefore future work will focus on the secondlevel algorithm, i.e. the resolution of superimposed MUAPs. Acknowledgments—The authors would like to thank David Hewson for his kind assistance with the language.
References BASMAJIAN, J. V., and DE LUCA, C. J. (1985): ‘Muscle alive. Their function revealed by electromyography’ (Williams & Wilkins, Baltimore, USA, 1985) CHRISTODOULOU, C. I., and PATTICHIS, C. S. (1999): ‘Unsupervised pattern recognition for the classification of EMG signals’, IEEE Trans. Biomed. Eng., 46, pp. 169–178 DUCHEˆNE, J., and HOGREL, J. Y. (2000): ‘A model of EMG Generation’, IEEE Trans. Biomed. Eng., 47, pp. 192–201 FANG, J., AGARWAL, G. C., and SHAHANI, B. T. (1999): ‘Decomposition of multiunit electromyographic signals’, IEEE Trans. Biomed. Eng., 46, pp. 685–696 HOGREL, J.-Y., and DUCHENE, J. (1999): ‘A SEMG-based system for clinical applications using laplacian electrodes’, in HERMENS, H. J., and FRERIKS, B. (Eds): ‘Future applications of surface electromyography, SENIAM 4’ (Roessingh Research and Development b.v., Enschede), pp. 172–177 IANI, C., STA˚LBERG, E., FALCK, B., and BISHOFF, C. (1994): ‘New approaches to motor unit potential analysis’, Ital. J. Neurol. Sci., 15, pp. 447–459 LEFEVER, R. S., and DE LUCA, C. J. (1982a): ‘A procedure for decomposing the myoelectric signal into its constituent action potentials-Part I: Technique, theory, and implementation’, IEEE Trans. Biomed. Eng., 29, pp. 149–157
Medical & Biological Engineering & Computing 2003, Vol. 41
LEFEVER, R. S., and DE LUCA, C. J. (1982b): ‘A procedure for decomposing the myoelectric signal into its constituent action potentials-Part II: Execution and test for accuracy’, IEEE Trans. Biomed. Eng., 29, pp. 158–164 MAMDANI, E. H., and ASSILIAN, S. (1975): ‘An experiment in linguistic synthesis with a fuzzy logic controller’, Int. J. ManMach. Stud., 7, pp. 1–13 MCGILL, K., CUMMINS, K. L., and DORFMAN, L. J. (1985): ‘Automatic decomposition of the clinical electromyogram’, IEEE Trans. Biomed. Eng., 32, pp. 470–477 RAU, G., and DISSELHORST-KLUG, C. (1997): ‘Principles of highspatial-resolution surface EMG (HSR-EMG): single motor unit detection and application in the diagnosis of neuromuscular disorders’, J. Electromyogr. Kinesiol., 7, pp. 233–239 REUCHER, H., RAU, G., and SILNY, J. (1987a): ‘Spatial filtering of noninvasive multielectrode EMG: Part I—Introduction to measuring technique and applications’, IEEE Trans. Biomed. Eng., 34, pp. 98–105 REUCHER, H., SILNY, J., and RAU, G. (1987b): ‘Spatial filtering of noninvasive multielectrode EMG: Part II—Filter performance in theory and modelling’, IEEE Trans. Biomed. Eng., 34, pp. 106–113 ROELEVELD, K., STEGEMAN, D. F., FALCK, B., and STA˚LBERG, E. V. (1997): ‘Motor unit size estimation: confrontation of surface EMG with macro EMG’, Electroencephalogr. Clin. Neurophysiol., 105, pp. 181–188 STASHUK, D. W., and QU, Y. (1996a): ‘Adaptive motor unit action potential clustering using shape and temporal information’, Med. Biol. Eng. Comput., 34, pp. 41–49 STASHUK, D. W., and QU, Y. (1996b): ‘Robust method for estimating motor unit firing-pattern statistics’, Med. Biol. Eng. Comput., 34, pp. 50–57 STASHUK, D. W. (1999): ‘Decomposition and quantitative analysis of clinical electromyographic signals’, Med. Eng. Phys., 21, pp. 389–404 STASHUK, D. W. (2001): ‘EMG signal decomposition: how can it be accomplished and used?’, J. Electromyogr. Kinesiol., 11, pp.151–173 TAKAGI, T., and SUGENO, M. (1985): ‘Fuzzy identification of systems and its application to modelling and control’, IEEE Trans. Syst. Man Cybern., 15, pp. 116–132 ZADEH, L. A. (1965): ‘Fuzzy sets’, Inf. Control, 8, pp. 338–353 ZADEH, L. A. (1978): ‘Fuzzy sets as a basis for a theory of possibility’, Fuzzy Set Syst., 1, pp. 3–28
Author’s biography ERIC CHAUVET received his BSc degree in Computer Engineering from the Department of Computer science in 1997 at the University of Technology of Compie`gne (France), where he is currently a PhD student in the Department of Biomedical Engineering. His research interests include signal modelling and analysis of bioelectric signals in particular surface EMG signals.
653
