Surface EMG decomposition requires an appropriate validation.

TO THE EDITOR: A major constraint in the analysis of motor unit activity in humans is that relatively few active motor units can be detected concurrently due to the high selectivity of needle electrodes commonly used for electromyographic (EMG) recordings. Furthermore, the accurate decomposition of intramuscular EMG signals is usually not possible at high muscle forces due to the many motor unit action potentials that overlap in time. De Luca and Hostage (2010) appear to have developed a noninvasive solution to these problems. They assert that it is possible to decompose surface EMG signals and detect with confidence the concurrent activity of 20–30 motor units at contractions up to the maximal force. Moreover, their decomposition method requires only four surface EMG channels. The main difficulty associated with processing surface EMG recordings is the greater number of overlapping action potentials than that in intramuscular EMG. For example, the second discharges of motor units #17 to #24 in their Fig. 10 occur at a similar instant in time and thus the action potentials of these eight motor units must partially overlap. Indeed, close inspection of Fig. 10 suggests that even more action potentials overlap at this point in time and at other times throughout the recording, as can also be appreciated from the interference surface EMG signal (see their Fig. 8). This is not surprising given the many action potentials that contribute to the surface EMG signal at each instant in time. Consequently, relatively few action potentials of each unit are detected as isolated in time from those of other units (Nawab et al. 2010). In principle, the overlapping action potentials can be disentangled by estimating the likelihood of each event in the multiple possible sets of matching templates. However, the global optimization of overlapping action potentials is a nondeterministic polynomial–type hard problem that cannot be solved by polynomial complexity algorithms (Ge et al. 2010). The maximal number of overlapping potentials that can be distinguished in each segment is practically limited by the size of the search space. A potential solution is to limit the search space to the most likely combinations and to include some constraints on the estimated discharge pattern (Nawab et al. 2010), but the likelihood of finding the correct solution decreases with the size of the search space (Ge et al. 2010). Consequently, the results achieved by De Luca and Hostage (2010) to decompose surface EMG signals are impressive. Such a remarkable achievement, however, requires an extremely convincing validation of the decomposition results. The results of the study by De Luca and Hostage (2010) were tested with an approach called the reconstruct-and-test procedure (Nawab et al. 2010), which they suggest is superior to the two-sensor method (Mambrito and De Luca 1984). The new validation approach involves generating a synthetic signal from the decomposed trains of action potentials and reapplying the same decomposition method to this synthetic signal to which noise is added with a power similar to that of the residual in the first decomposition (see their Fig. 9 and detailed description by Nawab et al. 2010). The accuracy in the identification of each train of action potentials is then estimated in the same way as in the two-sensor method; that is, as the percentage of action potentials that are identified by the two decompositions with respect to the total. We beg to differ that such an approach provides a valid demonstration of the accuracy of the decomposition algorithm. One of the problems with the reconstruct-and-test procedure is that missed discharges do not influence the index of accuracy. If the first decomposition does not detect a discharge time, the absent discharge will not be used to build the synthetic signal for the second decomposition. It is thus very likely that the second decomposition will not detect that missing discharge and thus will corroborate the error introduced in the first decomposition. Consequently, a disagreement in the results of the two decomposition methods is biased toward discharges detected by the first decomposition and not by the second, whereas there were no converse cases of discharges detected by the second decomposition and not by the first (their Fig. 10). This result differs from that obtained with the twosensor validation method, where either of the two decompositions can miss a discharge time with respect to the other. Contrary to the two-sensor method, the reconstruct-and-test procedure is biased in that the signal used in the second decomposition depends on the result of the first decomposition and may lead to an estimation of 100% accuracy for a train of action potentials, even when a substantial number of discharge times are not identified. Besides the failure to assess missed discharges, the reconstruct-and-test approach is not an appropriate validation method due to a more general issue. As reported by Nawab et al. (2010) and similar to many other decomposition methods, the decomposition algorithm used by De Luca and Hostage (2010) is data-driven, in that its output is identical when applied to identical signals. When the power of the residual signal from the original decomposition is low and the noise added to the synthesized signal is small, the original and synthesized signals will be relatively similar and thus the results of the two decompositions will be in close agreement, independent of the decomposition accuracy. As an extreme case, the accuracy index used by De Luca and Hostage will necessarily be equal to 100% for each train of action potentials when the residual signal is equal to zero. In practice, low power in the residual signal results in the algorithm being applied to two signals that are almost identical (as can be seen in Fig. 9 Address for reprint requests and other correspondence: D. Farina, Center for Sensory-Motor Interaction (SMI), Department of Health Science and Technology, Aalborg University, Aalborg, Denmark (E-mail: [email protected]). J Neurophysiol 105: 981–982, 2011. doi:10.1152/jn.00855.2010.