A critical review by

De Luca CJ and Kline JC

June 2015

Technology for decomposing the EMG signal into individual motor unit action potentials (MUAPs) was introduced by us three decades ago (LeFever and De Luca, 1982). The Delsys dEMG decomposition system is the fruition of three decades of research and development. It substantially outperforms other decomposition systems. It consists of:

Small non-invasive sensor
Sophisticated decomposition algorithm
Novel error correction algorithm
High-precision validation

Decomposition Sensor Technologies

Advantages of the Delsys Small dEMG Sensor

Our radically new non-invasive sensor is specially designed (De Luca et al, 2006) to acquire sEMG signals that can be decomposed into MUAPs by our dEMG algorithm.

The geometry and the dimensions of the dEMG sensor are specifically selected to detect four channels of differential sEMG signals of a quality that renders them particularly useful for the pragmatic execution of the dEMG algorithm. Its effectiveness is supported by several advantages:

  1. A simple, 5 x 5 mm, five-pin geometry ensures that the sensor is easy to apply and maintains long-lasting electrode contact throughout an experiment.
  2. The small-footprint is versatile for both small and large muscles.
  3. It acquires sEMG signals rich in motor unit action potential trains (MUAPTs).
  4. Its design is fine-tuned to detect distinct uncontaminated MUAPs embedded within sEMG signals essential for the proper working of the dEMG algorithm.

Other Decomposition Sensor Technologies

These are the multiple-electrode sensor arrays introduced by Monster and Chan (1980), applied clinically by Blok et al (2002) and Farina and Merletti (2004) and more recently used by Holobar et al (2012), Farina et al (2014), among others. They are limited because:

  1. They can be difficult to maintain consistent contact for all the large number (64 to 121) of electrodes “either due to bad electrode–skin contact or short circuit between two or more surface electrodes” according to Marateb et al (2011).
  2. The relatively large size makes them challenging to use with small muscles.
  3. The sEMG signals are recorded in a monopolar configuration, rendering them susceptible to movement artifacts and to noise. The resultant distortion that occurs to the MUAPs challenges the capability of any decomposition attempt.

Decomposition Algorithms

Advantage of the Delsys dEMG algorithm

Performance

  • Recover MUAP shapes and firing instances.
  • Yield as many as 50 MUAPTs (typically 20-30) active during contractions ranging from 5 to 80% of the maximal voluntary contraction.
  • Accuracy verified for each extracted MUAPT – on average 95%.

Our dEMG algorithm for decomposing sEMG signals is described in peer-reviewed publications by De Luca et al (2006) and Nawab et al (2010). It is an evolution of the algorithm first reported by LeFever and De Luca (1982). The dEMG algorithm is set apart from other approaches in the field because:

  1. It is grounded in fundamental principles of signal processing that are applied with no compromise.
  2. It uses template matching to track the changing shapes of MUAPs as they occur throughout the sEMG signal.
  3. It is agnostic to the underlying motor unit firing behavior.

Other Decomposition Algorithms

Performance

  • Recover only motor unit firing instances.
  • Yield up to 20 MUAPTs (typically 8-10) active during contractions ranging from 10 to 70% of the maximal voluntary contraction.
  • Accuracy is unknown for all MUAPTs.

These algorithms ignore the fundamental concepts upon which they are based. Consider the convolution kernel compensation (CKC) algorithm developed by Holobar and Zazula (2003, 2004, 2007) and promulgated by Farina, Merletti and Enoka (2014).Their approach is based on a signal processing technique called Independent Component Analysis (ICA). ICA was developed to identify independent sources in a mixture based on a relatively high number of observations. But in order for it to work in EMG decomposition, it is necessary to make assumptions about the characteristics of the EMG signal so that they obey the construct of ICA. In realistic physiological conditions these assumptions do not hold:

The firings of each motor unit must occur independently of the firings of all other MUAPTs contributing to the sEMG signal. But, during virtually all types of voluntary contractions there exist dozens of known and well-documented reports of synchronized firing instances by De Luca et al (1993), Winges and Santello (2004), Holtermann et al (2009), De Luca and Kline (2014); among others, and correlated fluctuations in firing rates, known as “common drive” by De Luca et al (1982), De Luca and Erim (1994), Laine et al (2013), Farina et al (2014) among others. Motor unit firings are not independent.

The action potentials of different motor units must remain stationary and not change throughout the duration of a contraction. But changes in the shape of the individual MUAP – reported by De Luca (1984), Juel (1988), Bertram et al (1995), Roy et al (2007), Fortune and Lowery (2009), among others – are known to occur in real sEMG signals as a function of contraction time or electrode movement. MUAPs are not stationary.

The number of different MUAPTs obtained from decomposition must always be less than or equal to the number of sEMG signals recorded. Algorithms based on this assumption are restricted to using cumbersome multi-electrode arrays subject to shortcomings outlined in section 1 above.

These ICA precepts are well documented in classical signal processing books (for example Duda et al, 2000).

We are not alone in expressing these concerns. Chen and Zhou (2015) have pointed out these limiting issues. Even Holobar and Zazula (2003, 2004, 2007) acknowledged these limitations when developing the CKC algorithm; although they later contrived artificial models in an attempt to show that the CKC could overcome these foundational assumptions. But, as we have pointed out in De Luca and Nawab (2011) and Kline and De Luca (2014): fabricated models reported in the literature do not capture all the physiological characteristics of the real sEMG signal. As such, the cogency of the CKC decomposition algorithm remains untested.

Decomposition Error Reduction

Advantage of the Delsys Decomposition Error Reduction Algorithm

All decomposition algorithms are subject to imperfection; even those relying on visual identification of individual MUAPs by a human operator. We have developed a new process by which these decomposition errors can be identified and mitigated to improve the decomposition result (Kline and De Luca, 2014). The basic concept of the error-reduction algorithm consists of decomposing the recorded sEMG signal into multiple estimates of the constituent MUAPTs and combining these estimates to derive a new estimate with fewer and smaller decomposition errors.

Error Reduction Used by Others

None.

Verification of Decomposition Accuracy

Advantage of the Delsys Verification of Decomposition Accuracy

Our Decompose, Synthesize, Decompose, Compare (DSDC) validation (Nawab et al, 2010; De Luca and Contessa, 2012) overcomes the shortcomings of other tests. It uses a physiologically realistic signal synthesized from MUAPTs obtained from the decomposition of a real sEMG signal. By adding Gaussian noise to the synthesized signal we create a unique signal that is of the same class as the real sEMG signal. By decomposing the synthesized signal and comparing the extracted MUAPTs with those known within the synthesized signal we are able to obtain the accuracy and location error of decomposition for each MUAPT. The efficacy of our DSDC validation is supported by several advantages:

  1. It validates the accuracy of each of the MUAPTs extracted from the sEMG signal.
  2. It makes no assumption about the characteristics of the MUAP shapes or variation of these shapes throughout the sEMG signal.
  3. It is agnostic to the underlying firing behavior of individual motor units.

Other Verifications of Decomposition Accuracy

Validation using mathematically synthesized signals

This approach uses a signal synthesized from artificially generated MUAPs and firing instances that are designated as the “truth”. The synthesized signal is decomposed and the accuracy is evaluated by comparing the decomposed data with the truth data.

But, the use of a mathematically synthesized signal is a generic test used under artificial circumstances. It requires the presumption that the accuracy results obtained under artificial conditions provide a faithful representation of the decomposition accuracy of a real sEMG signal. This is a difficult point to establish when the mathematically synthesized signal does not contain some of the complexities found in real physiological signals. For example current mathematical models use fabricated MUAPs that are unrealistically smooth because they are generated by models where the depolarization is surrounded by a homogeneous environment as is evidenced by the volume conductor model proposed by Merletti et al (1999), Farina and Merletti (2001) and Lowery et al (2002). The artificial smooth shapes of the MUAPs do not capture and represent the richer irregular details of the shapes of real MUAPs nor do they consider the changes in MUAP shape that occur in normal physiological conditions. In addition, currentmathematical models use fabricated motor unit firing instances that are either constructed artificially, as was done by Holobar and Zazula (2007) and Holobar et al (2012), or are derived from a generic model, such as that reported by Fuglevand et al (1993). These approaches fail to integrate known and documented physiological firing behavior such as synchronization and common drive. Yet, they have been used extensively by Holobar and Zazula (2004), Holobar et al (2009) and Holobar et al (2014) as the primary validation of the CKC decomposition algorithm. Because mathematically synthesized models do not directly assess the decomposition of a real sEMG signal the accuracy of individual extracted MUAPTs cannot be measured and remains unknown.

The test was developed by our own group more than 3 decades ago (Mambrito and De Luca, 1984), and has been used by Holobar et al (2009), Marateb et al (2011), Holobar et al (2011), and Farina, Merletti and Enoka (2014) to justify accurate performance of the CKC decomposition algorithm. The test evaluates the decomposition accuracy by comparing common MUAPTs extracted from two EMG signals recorded simultaneously from two sensors arranged in near proximity.

Although capable of estimating the accuracy of a select few MUAPTs, the two-source test falls far short of providing a comprehensive validation. Only a small fraction of MUAPTs common across two signals can be assessed leaving the accuracy of the remaining MUAPTs untested.

Validation tests developed by McGill and Marateb (2011) and separately by Parsaei and Stashuk (2013) evaluate decomposition accuracy based on statistical expectations and/or assumptions of motor unit firing behavior.

This is a generic validation approach. It applies assumptions – such as stationary MUAP shapes and independent motor unit firing instances – that do not hold under normal physiological conditions (see detailed explanation in Section 2). Furthermore, the validation is limited to a small class of signals that “should not be too complex” according to McGill and Marateb (2011). But it is in complex portions of the sEMG signal where decomposition errors are most likely to occur. Because this test is unable to function in these regions, it is incapable of providing a useful assessment of the accuracy of the extracted MUAPTs.

Recently, Holobar et al (2014) reported that the pulse-to-noise ratio – a measure of the MUAP signal to noise ratio – can be used to assess the accuracy of decomposition. They used the two-source test to measure the sensitivity (a measure of accuracy) of extracted MUAPTs and plotted the values as a function of the MUAP signal to noise ratio. They claimed a relationship between the two variables indicated the pulse-to-noise ratio was a reliable indicator of decomposition accuracy.

But, the MUAP signal to noise ratio does not measure actual accuracy because it ignores the incidence of identification and location errors that occur throughout the decomposition. Furthermore, the correlations between their measure of accuracy and the pulse-to-noise ratio had R2 values as low as 0.41. In other words, in some contractions, more than 50% of the variance in the accuracy data could not be explained by the pulse-to-noise ratio. The actual accuracy for extracting each MUAPT remains unknown using the pulse-to-noise ratio.

Contemporary Discourse in the Literature

In a recent letter to the editor of the Journal of Applied Physiology (De Luca et al, 2015) we challenged the validation recommendations proposed by Farina, Merletti and Enoka (2014) and gave the authors an opportunity to respond to the following concerns. They provided no responses.

Point 1 Previously Farina and Enoka (2011) claimed that decomposition of synthesized mathematically-generated signals is the best way to validate sEMG decomposition algorithms. But this approach does not test the accuracy of decomposing a real sEMG signal. As examples consider models developed by Farina and Merletti (2001) that ignore known physiological characteristics of MUAP shapes, firing instances or dependent firing behavior such as synchronization and common drive. Yet, these models have been used as the primary validation of the CKC decomposition algorithm. In a previous letter we (De Luca and Nawab, 2011) pointed out these limitations. Now Farina, Merletti and Enoka (2014) have reversed their position and admit that synthesized mathematically generated signals are “more limited than an experimental validation”. Their reversal was made clear in our letter (De Luca et al, 2015).
Point 2 As an alternative validation Farina, Merletti and Enoka (2014) now propose the “only current reliable approach to assess the accuracy of a surface EMG decomposition algorithm” is the two-source test. As discussed above, this test was developed and first used by us more than 4 decades ago. Being well aware of its strengths and limitations we expressed the primary drawback of the two-source test in our letter (De Luca et al, 2015).
Point 3 When Holobar et al (2014) used the two-source test to validate the CKC decomposition algorithm, they could only validate on average 0.7 MUAPTs from each contraction and reported some accuracy values less than 60%. How such low accuracy values from such a small percentage of MUAPTs establishes the validity of the CKC algorithm for sEMG decomposition remains confounding.
Point 4 To avoid the drawbacks of these other approaches we developed the decompose-synthesize-decompose-compare (DSDC) method to validate our dEMG algorithm. Farina, Merletti and Enoka (2014) claimed that our method is flawed and does not provide a reliable assessment of decomposition accuracy. We pointed out that when the DSDC validation was compared with the fraction of MUAPTs obtained using their favored two-source test, both validations yielded 95% average accuracy for our dEMG algorithm across different subjects, muscles and force levels. Only the DSDC validation had the advantage of evaluating the accuracy of the comprehensive set of MUAPTs, far surpassing the select few assessed by the two-source test.

References

  1. Bertram MF, Nishida T, Minieka MM, Janssen I and Levy CE. Effects of temperature on motor unit action potentials during isometric contraction. Muscle & Nerve 18:1443-1446, 1995.
  2. Blok JH, Van Dijk JP, Drost G, Zwarts MJ & Stegeman DF. A high-density multichannel surface electromyography system for the characterization of single motor units. Review of scientific instruments 73:1887-1897, 2002.
  3. Chen M and Zhou P. A Novel Framework Based on FastICA for High Density Surface EMG Decomposition. IEEE Transactions on Neural Systems and Rehabilitation Engineering99:1-11, 2015.
  4. De Luca CJ. Myoelectric manifestations of localized muscular fatigue. CRC Critical Reviews in Biomedical Engineering 11:251-279, 1984.
  5. De Luca CJ, LeFever RS, McCue MP and Xenakis AP. Control scheme governing concurrently active human motor units during voluntary contractions. Journal of Physiology329:129-142, 1982.
  6. De Luca CJ, Roy AM and Erim Z.Synchronization of motor unit firings in human muscles. Journal of Neurophysiology70:2010-2023, 1993.
  7. De Luca CJ and Erim Z. Common drive of motor units in regulation of muscle force. Trends in Neurosciences 17:299–305, 1994.
  8. De Luca CJ, Adam A, Wotiz R, Gilmore LD and Nawab SH. Decomposition of surface EMG signals. Journal of Neurophysiology 96:1646-1657, 2006.
  9. De Luca CJ and Nawab SH. Reply to Farina and Enoka: The reconstruct-and-test approach is the most appropriate validation for surface EMG signal decomposition to date. Journal of Neurophysiology 105:983-984, 2011.
  10. De Luca CJ and Contessa P. Hierarchical control of motor units in voluntary contraction. Journal of Neurophysiology 107:178-195, 2012.
  11. De Luca CJ and Kline JC. Statistically rigorous calculations do not support Common Input and Long-Term synchronization of motor unit firings. Journal of Neurophysiology112:2729-2744, 2014.
  12. De Luca CJ, Nawab SH and Kline JC. Clarification of methods used to validate surface EMG decomposition algorithms as described by Farina et al. (2014). Journal of Applied Physiology118:1084-1084, 2015.
  13. Duda RO, Hart PE and Stork DG. Pattern Classification, 2nd Edition, Wiley, 2000.
  14. Farina D and Merletti R. A novel approach for precise simulation of the EMG signal detected by surface electrodes. IEEE Transactions on Biomedical Engineering 48:637-646, 2001.
  15. Farina D and Merletti R. Estimation of average muscle fiber conduction velocity from two-dimensional surface EMG recordings. Journal of Neuroscience Methods 134:199-208, 2004.
  16. Farina D and Enoka RM. Surface EMG decomposition requires an appropriate validation. Journal of Neurophysiology 105:981-982, 2011.
  17. Farina D, Rehbaum H, Holobar A, Vujaklija I, Jiang N, Hofer C Salminger S, van Vliet HW and Aszmann O. Noninvasive, accurate assessment of the behavior of representative populations of motor units in targeted reinnervated muscles. IEEE Transactions onNeural Systems and Rehabilitation Engineering 22:810-819, 2014.
  18. Farina D, Merletti R and Enoka RM. The extraction of neural strategies from the surface EMG: an update. Journal of Applied Physiology 117:1215–1230, 2014.
  19. Fortune E and Lowery MM. Effect of extracellular potassium accumulation on muscle fiber conduction velocity: a simulation study. Annals of Biomedical Engineering37:2105-2117, 2009.
  20. Fuglevand AJ, Winter DA and Patla AE.Models of recruitment and rate coding organization in motor-unit pools. Journal of Neurophysiology 70:2470–2488, 1993.
  21. Holobar A and Zazula D. Surface EMG decomposition using a novel approach for blind source separation. Informatica Medica Slovenica 8:2–14, 2003.
  22. Holobar A and Zazula D. Correlation-based decomposition of surface electromyograms at low contraction forces. Medical & Biological Engineering & Computing 42:487–495, 2004.
  1. Holobar A and Zazula D. Multichannel blind source separation using convolution kernel compensation. IEEE Transactions on Signal Processing 55:4487-4496, 2007.
  2. Holobar A, Minetto MA, Botter A and Farina D. Identification of motor unit discharge patterns from high-density surface EMG during high contraction levels. 5th European Conference of the International Federation for Medical and Biological Engineering Proceedings37:1165-1168, 2011.
  3. Holobar A, Glaser V, Gallego JA, Dideriksen JL and Farina D. Non-invasive characterization of motor unit behaviour in pathological tremor. Journal of Neural Engineering 9:1-13 2012.
  4. Holobar A, Farina D, Gazzoni M, Merletti R and Zazula D. Estimating motor unit discharge patterns from high-density surface electromyogram. Clinical Neurophysiology120:551-562, 2009.
  5. Holobar A, Minetto MA and Farina D. Accurate identification of motor unit discharge patterns from high-density surface EMG and validation with a novel signal-based performance metric. Journal of Neural Engineering 11:1-11, 2014.
  6. Holtermann A, Grönlund C, Karlsson JS and Roeleveld K. Motor unit synchronization during fatigue: Described with a novel sEMG method based on large motor unit samples. The Journal of Electromyography and Kinesiology 19: 232-241, 2009.
  7. Juel C. Muscle action potential propagation velocity changes during activity. Muscle & Nerve11:714-719, 1988.
  8. Kline JC and De Luca CJ. Error Reduction in EMG Signal Decomposition. Journal of Neurophysiology 112:2718-2728, 2014.
  9. Laine CM, Negro F and Farina D. Neural correlates of task-related changes in physiological tremor. Journal of Neurophysiology 110:170-176, 2013.
  10. LeFever RS and De Luca CJ. A procedure for decomposing the myoelectric signal into its constituent action potentials. Part I. Technique, theory and implementation. IEEE Transactions on Biomedical Engineering 29:149-157, 1982.
  11. Lowery MM, Stoykov NS, Taflove A and Kuiken TA. A multiple-layer finite-element model of the surface EMG signal. IEEE Transactions on Biomedical Engineering 49:446-454, 2002.
  12. Mambrito B and De Luca CJ. A technique for the detection, decomposition and analysis of the EMG signal. Electroencephalography and Clinical Neurophysiology 58:175-188, 1984.
  13. Marateb HR, McGill KC, Holobar A, Lateva ZC, Mansourian M and Merletti R. Accuracy assessment of CKC high-density surface EMG decomposition in biceps femoris muscle. Journal of Neural Engineering 8:1-11, 2011.
  14. McGill KC and Marateb HR. Rigorous a posteriori assessment of accuracy in EMG decomposition. IEEE Transactions on Neural Systems and Rehabilitation Engineering 19:54-63, 2011.
  15. Merletti R, Lo Conte L, Avignone E and Guglielminotti P. Modeling of surface myoelectric signals. I. model implementation. IEEE Transactions on Biomedical Engineering46:810-820, 1999.
  16. Monster AW and Chan H. Surface electromyogram potentials of motor units; relationship between potential size and unit location in a large human skeletal muscle. Experimental Neurology67:280-297, 1980.
  17. Nawab SH, Chang SS and De Luca CJ. High-yield decomposition of surface EMG signals. Clinical Neurophysiology 121:1602-1615, 2010.
  18. Parsaei H and Stashuk DW. EMG signal decomposition using motor unit potential train validity. IEEE Transactions on Neural Systems and Rehabilitation Engineering 21:265-274, 2013.
  19. Roy, SH, De Luca G, Cheng MS, Johansson A, Gilmore LD and De Luca CJ. Electro-mechanical stability of surface EMG sensors. Medical & Biological Engineering & Computing45:447-457, 2007.
  20. Winges SA & Santello M. Common input to motor units of digit flexors during multi-digit grasping. Journal of Neurophysiology 92: 3210-3220, 2004.