TL;DR
This paper introduces a novel sequential estimation algorithm for intramuscular EMG signals, aiming to enable real-time control of upper-limb prostheses. By modeling EMG signals as a sum of filtered spike trains from individual motor units, the method efficiently estimates motor unit firing rates, paving the way for more intuitive and responsive prosthetic control.
What Problem Does This Paper Address?
Controlling advanced prosthetic limbs requires understanding the user's intended movements. Electromyography (EMG) signals, which reflect the electrical activity of muscles, are a promising source of this control information. While surface EMG has been used, intramuscular EMG, recorded directly within muscles, offers richer data about individual motor unit activity. However, processing these signals in real-time for prosthesis control is computationally challenging due to the complexity of decomposing and interpreting the superimposed signals from multiple motor units. This paper tackles the problem of efficiently and accurately estimating the firing rates of these individual motor units from intramuscular EMG data in a sequential manner, suitable for real-time applications.
Methods
The authors propose a Bayesian filtering approach to sequentially estimate motor unit firing rates. They model the intramuscular EMG signal ($y[n]$) as a linear sum of filtered spike trains ($u_i[n]$) from $M$ motor units, each with a specific impulse response ($h_i$), plus additive Gaussian noise ($w[n]$):
$y[n] = \sum_{i=1}^{M} (h_i * u_i)[n] + w[n]$
Each $u_i[n]$ is treated as an independent Bernoulli-Gaussian process with an unknown firing rate $q_i$. The core of their method involves:
- Path Enumeration: At each time step, they consider all possible combinations of motor unit firings (spike trains) up to that point.
- Posterior Probability Calculation: They compute the posterior probability of each possible path (sequence of firings) given the observed EMG data, using Bayes' theorem. This involves combining the prior probability of a path with the likelihood of observing the current EMG signal given that path.
- Path Pruning: To manage computational complexity, they retain only the $K$ most probable paths at each step.
- Firing Rate Estimation: The estimated firing rate for each motor unit is calculated as the weighted average of the firing rates across the selected paths, where the weights are the normalized posterior probabilities of those paths.
- Tracking Extension: A modified algorithm is introduced to handle time-varying firing rates by incorporating a forgetting factor, allowing the system to adapt to changes in motor unit activity.
Datasets
The paper evaluates their method on two types of data:
- Simulated Data: Generated using the proposed model (equation 1), with varying numbers of motor units (1-6), signal-to-noise ratios (15-30 dB), and spike train probabilities.
- Experimental Data: Intramuscular EMG signals recorded from the biceps brachii muscle of five healthy men during isometric contractions at 5% or 10% of their maximal voluntary contraction (MVC) force. These signals were pre-processed (amplified, band-pass filtered) and sampled at 10 kHz. The experimental data was compared against results obtained from manual decomposition by experts using the EMGLAB tool.
Key Insights and Ideas
- Bridging the Gap: The paper effectively bridges the gap between complex intramuscular EMG signal decomposition and the real-time demands of prosthetic control by introducing a sequential, Bayesian approach.
- Model-Based Estimation: By explicitly modeling the EMG signal as a sum of filtered spike trains, the method allows for a principled estimation of underlying neural control signals (firing rates).
- Computational Feasibility: The path pruning strategy is crucial for making the otherwise computationally intractable problem of enumerating all possible spike trains feasible for real-time applications.
- Adaptability: The inclusion of a forgetting factor in the tracking algorithm is a key insight for handling the inherent variability and time-varying nature of biological signals like EMG.
Key Results
- Simulated Data: The algorithm demonstrated good performance on simulated data, with estimated parameters closely tracking the true values after a short initial period. The use of a forgetting factor corresponding to a 2-second window length provided smooth results while allowing for tracking.
- Experimental Data: When applied to experimental data, the proposed method showed promising results, with estimated firing rates generally following the trends identified by expert operators. However, some discrepancies were observed for specific motor units, which the authors attribute to limitations in their simplified model (e.g., time-invariant MUAP shapes and potential interference).
Limitations and Future Work
The authors acknowledge several limitations:
- Simplified Model: The current model assumes time-invariant MUAP shapes and does not account for the refractory period between motor unit impulses.
- Manual Parameter Estimation: For experimental data, the noise variance, number of firing motoneurons, and MUAP shapes were estimated manually or assumed known.
- Correlated Activities: The model treats motor unit activities as independent, which is not entirely true in biological systems.
Future work will focus on:
- Online MUAP Shape Estimation: Developing methods to automatically estimate the shape of motor unit action potentials in real-time.
- Automatic Number of Firing Motoneurons Estimation: Enabling the algorithm to determine the number of active motor units dynamically.
- Incorporating Refractory Periods and Correlations: Enhancing the model to account for more realistic physiological constraints, such as the refractory period and correlations between motor unit activities.
Summary updated