Research

Objectives

To develop new tools for forward modelling in EMG. We will exploit the existing knowledge of Cronos (previously Athena) in building such tools (e.g. OpenMEEG) and see how they can be adapted to this new domain. One of the anticipated challenge will be to accurately extract anatomical information from MRI, such as muscle and bone surfaces. Another challenge will be dealing with volume deformations that occur during arm movements.

To develop new algorithms for the estimation of motor unit spike trains from surface EMG community recordings that exploit knowledge of the volume conductor. In particular, we will focus on the adaption of methods developed by the EEG such as the minimum norm estimate and maximum entropy on the mean.

To quantify the performance of these new algorithms by testing them on real EMG data of the forearm during hand movements. Here, we will exploit the availability of intramuscular recording that will provide ground truth data. Of course, intramuscular recordings cannot be acquired everywhere and one of the challenges will be to perform the validation using this partial data.

To validate EEG modelling approach on EMG data. One of the main challenges in EEG forward and inverse problems is the validation of the algorithms and approaches. Because they are extremely invasive, intracortical recordings associated to EEG data are rarely available. This makes it difficult to validate both forward models and inverse methods using real data, as there is no ground truth. In EMG, the situation is a different because intramuscular recordings are minimally invasive and therefore more easily accessible. This presents a unique opportunity to validate EEG algorithms on EMG data because of the identical physics supporting both modalities. Of course, results will have to be interpreted with care, as there are still some modelling differences, e.g. in EEG the bone sits between
the sources and the sensors instead of beneath in EMG.

Results

Year 1

In line with our first objective, which is to develop new tools for forward modelling in EMG, the associated team has proposed new methods to model EMG signals.

The first contribution is the development of BioMime, a conditional generative neural network trained adversarially to generate motor unit activation potential waveforms under a wide variety of volume conductor parameter. BioMime can be used to mimic the dynamic changes of a biophysical system that matches the mechanical movements of a musculoskeletal mode as illustrated below:

Our results have been published in IEEE Transactions on Neural Networks and Learning Systems (https://ieeexplore.ieee.org/document/10636282) and the author version of the manuscript is available on Arxiv ( https://arxiv.org/abs/2211.01856).

The second contribution is the development of NeuroMotion, an open-source simulator that provides a full-spectrum synthesis of EMG signals during voluntary movements. It builds upon BioMime to generate EMG signals from a movement descriptor. NeuroMotion is comprised of three modules. The first module is an upper-limb MSK model with OpenSim API to estimate the muscle fibre lengths and muscle activations during movements. The second module is BioMime, a deep neural network-based EMG generator that receives nonstationary physiological parameter inputs, such as muscle fibre lengths, and efficiently outputs motor unit action potentials (MUAPs). The third module is a motor unit pool model that transforms the muscle activations into discharge timings of motor units. The discharge timings are convolved with the output of BioMime to simulate EMG signals during the movement.

Our results have been published in PLOS Computational Biology (https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012257) and the author version of the manuscript is available on bioRxiv (https://www.biorxiv.org/content/10.1101/2023.10.05.560588).

Year 2

As planned for the second year, the team has also made progress on its second objective: developing novel algorithms to estimate motor unit spike trains from surface EMG. Two different approaches are pursued in parallel: electrode location optimization and physics-informed inverse problems.

Electrode location optimization: The first approach is based on the idea that the electrode configuration is not optimized and leads to variations in sensitivity across populations. For example, the number of identified motor neurons identified is typically larger in males than in females. We also observe variations in sensitivity across target muscle groups within a single individual. Using the forward models developed during the first year, we will generate highly realistic EMG signals representing a large population of individuals with varying anatomical features. The large amount of data generated will allow us to optimize electrode design globally, for specific populations (females/males), or for specific muscle architectures (e.g. pennate vs fusiform). The strength of this approach is to leverage highly realistic simulations to optimize over a large population whose EMG signals would be unfeasible to acquire in vivo (potentially thousands of individuals). Experiments are ongoing and we expect initial results to be published shortly.

Physics-informed inverse problems: The second approach leverages volume conductor knowledge to enhance the estimation of motor unit spike trains. Using anatomical MRI data of the forearm —acquired in various positions during the first year with optimized sequences — the team has constructed physically realistic forward models for EMG in two subjects. We are now investigating the theoretical aspects of the problem, building on our expertise in electroencephalography (EEG). One of the challenges we are currently investigating is the inclusion of the known temporal components of EMG, which are typically available in EEG.

Year 3

In progress!