Recursive Decomposition of Electromyographic Signals With a Varying Number of Active Sources: Bayesian Modeling and Filtering

Tianyi Yu, Konstantin Akhmadeev, Éric Le Carpentier, Yannick Aoustin, Raphaël Gross, Yann Péréon, Dario FarinaIEEE Trans. Biomedical Engineering 2019 · DOI

TL;DR

This paper introduces a novel algorithm for decomposing electromyographic (EMG) signals, which are electrical signals from muscles. The key innovation is its ability to accurately track changes in the number of active muscle fibers over time, a significant improvement over previous methods that assumed a constant number. This makes the algorithm more robust and applicable to real-world scenarios, paving the way for advanced human-machine interfaces.

What Problem Does This Paper Address?

Electromyography (EMG) signals are generated by the electrical activity of motor units (MUs) within muscles. Decomposing these signals means identifying the individual firing times of these MUs and their associated electrical waveforms (Motor Unit Action Potentials or MUAPs). This decomposition provides crucial insights into neural control of movement and is valuable for diagnosing neuromuscular disorders and developing advanced prosthetics.

A major challenge in EMG decomposition is that the number of active MUs can change dynamically during muscle contraction (recruitment and derecruitment). Previous algorithms often struggled with this variability, assuming a fixed number of active MUs. This assumption led to errors, particularly during periods of changing muscle force, causing the algorithms to diverge or miss important events. This paper tackles this limitation by developing a method that can adapt to a varying number of active MUs.

Methods

The core of the proposed method is a recursive Bayesian filtering framework built upon a Hidden Markov Model (HMM). Here's a breakdown of the key components:

  • Hidden Markov Model (HMM): The iEMG signal is modeled as a system with hidden states. These states include:
  • Activation Scenario (S[n]): This represents which motor units are active at a given time n, and the time elapsed since their last spike.
  • MUAP Shapes (H[n]): The electrical waveform generated by each motor unit.
  • Inter-Spike Law Parameters (Θ[n]): Parameters governing the timing of spikes for each motor unit, modeled using a discrete Weibull distribution.
  • Dynamic Recruitment Model: This is the paper's main contribution. It explicitly models the recruitment and derecruitment of motor units. A motor unit is considered derecruited when the time since its last spike exceeds a threshold, and new units are recruited with a certain probability. This is a significant departure from prior work that assumed a constant number of active units.
  • Bayesian Filtering: The algorithm uses a recursive Bayesian filter to estimate the hidden states (activation, MUAP shapes, and spike timing parameters) from the observed iEMG signal. This involves:
  • Prediction: Estimating the state at the next time step based on the current state.
  • Update: Correcting the prediction using the new observed data.
  • Recursive Maximum Likelihood (RML) Estimation: Used to estimate the parameters of the inter-spike law (Weibull distribution).
  • Kalman Filter: Employed for estimating the MUAP shapes and their associated uncertainties.
  • Adaptation to Non-Stationarity: The algorithm incorporates sliding window approaches and adaptive coefficients to handle changes in MUAP shapes and noise variance over time.

Datasets

The algorithm was evaluated using two types of datasets:

  • Simulated Signals: Generated using the proposed HMM framework. These signals were created with varying numbers of active MUs (6, 8, and 10) and different Signal-to-Noise Ratios (SNR) (10 dB). MUAP shapes were derived from experimental data.
  • Experimental Signals: Acquired from two muscles:
  • Tibialis Anterior (TA): Using fine-wire electrodes during an isometric force tracking task.
  • Abductor Digiti Minimi (ADM): Using needle electrodes during different levels of finger abduction.

The paper does not introduce a new benchmark dataset but rather uses existing experimental data and simulated data based on a realistic model.

Key Insights and Ideas

  • Dynamic Recruitment is Crucial: The paper strongly demonstrates that accurately modeling the dynamic recruitment and derecruitment of motor units is essential for robust EMG decomposition, especially in real-world applications where muscle activity is rarely static.
  • Bayesian Filtering for Adaptability: The recursive Bayesian filtering approach, combined with an HMM, provides a powerful framework for simultaneously estimating multiple parameters (spike times, MUAP shapes, noise) and adapting to changes in the signal over time.
  • Weibull Distribution for ISI: The use of the discrete Weibull distribution for inter-spike intervals (ISIs) is a well-established choice that captures the rhythmic nature and refractory period of motor unit firing.
  • Parallelization Potential: The algorithm's structure, particularly the path selection in the HMM, lends itself well to parallel computation, which is key for achieving real-time performance.

Key Results

The algorithm demonstrated strong performance on both simulated and experimental data:

  • High Decomposition Accuracy:
    • For simulated signals, decomposition accuracy (Sens. and Pred.) exceeded 90% for up to 10 MUs.
  • For experimental signals, global sensitivity and predictivity were consistently above 90%.
  • Successful Tracking of Recruitment: The algorithm successfully tracked changes in MU recruitment and derecruitment, as evidenced by comparisons with reference decompositions.
  • Robustness to Superposition: The method effectively handled complex superpositions of MUAPs, even with up to 5 MUs overlapping simultaneously in some cases.
  • Adaptation to MUAP Changes: The algorithm showed the ability to correct initial rough estimates of MUAP shapes and track their gradual changes over time, a critical feature for real-world signals. For instance, an experiment where Kalman filtering was "switched off" resulted in significantly lower performance, highlighting the importance of adaptive MUAP estimation.

Limitations and Future Work

  • Computational Complexity: The paper acknowledges that the computational complexity can be high, especially when a large number of paths (npath) need to be considered, which occurs with a high number of active MUs. However, they suggest that parallel implementation can mitigate this.
  • Dependence on Initial MUAP Estimates: While the algorithm can correct initial estimates, it still relies on some form of initial MUAP extraction. The authors suggest that automated MUAP clustering techniques could be used to address this.
  • Real-time Implementation: The paper lays the groundwork for real-time applications, but the current implementation's computational demands might still be a barrier for some high-performance real-time scenarios. Future work will focus on optimizing this for real-time use in human-machine interfaces.
  • Further Validation: While tested on two muscles, further validation on a wider range of muscles and conditions would be beneficial.

Summary updated

Abstract

The proposed method substantially broadens the domains of applicability of the algorithm.
BibTeX
@article{yu2019recursive,
  author = {Tianyi Yu and Konstantin Akhmadeev and Éric Le Carpentier and Yannick Aoustin and Raphaël Gross and Yann Péréon and Dario Farina},
  title = {{Recursive Decomposition of Electromyographic Signals With a Varying Number of Active Sources: Bayesian Modeling and Filtering}},
  journal = {IEEE Transactions on Biomedical Engineering},
  year = {2019},
  doi = {10.1109/tbme.2019.2914966},
}
Figure