Les commentaires sont clos.

Master thesis: An “on the fly” EMG decomposition

14 Octobre 2022

Catégorie : Stagiaire

Laboratory: Laboratoire des Sciences du Numérique de Nantes, Ecole Centrale de Nantes, Nantes Université. Research teams ReV (Robotique et Vivant) and SIMS (Signal, Image et Son).

Context: Electromyography (EMG) is used in routine in clinical practice Pereon (2015), Gallard 2020. Electrodes are inserted into the muscle (intramuscular EMG, iEMG) or placed on the skin (surface EMG, sEMG) to record the muscular electrical activity. The measured EMG signal is a sum of elementary contributions. Each contribution is a wavelet train produced by a motor unit (MU) in the electrode vicinity. A MU corresponds to a spinal moto-neuron (MN) and the muscle fibers it innervates Heckman and Enoka (2012 ), and the wavelet is called the Motor Unit Action Potential (MUAP). A variation of muscle activation level produces a variation of the number of active MUs in a process called ”spatial recruitment”, and a variation of the discharge rate (that is the number of MUAPs per time unit) of the active MUs, called ”temporal recruitment”. EMG decomposition jointly estimates motor unit action potential (MUAP) waveforms and discharge rate. This information can be applied by a medical practitioner for the diagnosis to detect neuropathy or myopathy diseases.

Motivations and general objectives:

The main conjecture of this work is that the code composed of the number of activated MUs, the associated MUPs shapes and firing rates, provides a reliable signature of the iEMG signal, which can be retrieved from EMG decomposition algorithms, and can be used for recognition purposes. A single channel decomposition algorithm has been developed by LS2N. This decomposition algorithm is based on a Hidden Markov Model of a sampled EMG. At each time index, a MN can produce a pulse or not, with a probability based on a probabilistic model of the rough periodicity of the pulse train. Figure 1 shows an extract of the experimental signal decomposition. This leads to an exponential number of activation scenarios, among which the scenario with the highest posterior probability is the estimated one. Heuristic rules of scenarios pruning have been proposed, and the intrinsically parallel structure of the algorithm leads to a GPU implementation [Yu et al. 2020a, 2020b], which can currently decompose up to 10 sources. To decompose a greater number of sources, a rigorous mathematical way to discard unnecessary scenarios is needed, which can prune a whole set of scenarios without the calculation of the posterior probability of each of them. This pruning way will be based on the recursive implementation of operation research techniques such as branch-­‐and-­‐bound optimization, currently a research field at LS2N [Ben Mhenni et al. 2021]. The final aim is to obtain an “on the fly” decomposition, particularly attractive for the practitioner.


Sébastien Bourguignon,