Task-discriminative space-by-time factorization of muscle activity

Ioannis Delis, Stefano Panzeri, Thierry Pozzo, Bastien BerretFrontiers in Human Neuroscience 2015 · DOI

Abstract

Movement generation has been hypothesized to rely on a modular organization of muscle activity. Crucial to this hypothesis is the ability to perform reliably a variety of motor tasks by recruiting a limited set of modules and combining them in a task-dependent manner. Thus far, existing algorithms that extract putative modules of muscle activations, such as Non-negative Matrix Factorization (NMF), identify modular decompositions that maximize the reconstruction of the recorded EMG data. Typically, the functional role of the decompositions, i.e., task accomplishment, is only assessed a posteriori. However, as motor actions are defined in task space, we suggest that motor modules should be computed in task space too. In this study, we propose a new module extraction algorithm, named DsNM3F, that uses task information during the module identification process. DsNM3F extends our previous space-by-time decomposition method (the so-called sNM3F algorithm, which could assess task performance only after having computed modules) to identify modules gauging between two complementary objectives: reconstruction of the original data and reliable discrimination of the performed tasks. We show that DsNM3F recovers the task dependence of module activations more accurately than sNM3F. We also apply it to electromyographic signals recorded during performance of a variety of arm pointing tasks and identify spatial and temporal modules of muscle activity that are highly consistent with previous studies. DsNM3F achieves perfect task categorization without significant loss in data approximation when task information is available and generalizes as well as sNM3F when applied to new data. These findings suggest that the space-by-time decomposition of muscle activity finds robust task-discriminating modular representations of muscle activity and that the insertion of task discrimination objectives is useful for describing the task modulation of module recruitment.

Introduction: The Need for Task-Aware Muscle Synergies

The central nervous system (CNS) likely simplifies movement control by recruiting "modules"—groups of muscles working together in specific patterns—rather than controlling every muscle fiber individually. Traditionally, researchers use Non-negative Matrix Factorization (NMF) to decompose electromyographic (EMG) signals into spatial and temporal components. However, standard NMF focuses solely on reconstructing raw data as accurately as possible, often ignoring the functional context of the movement. This can lead to models that overfit to task-irrelevant noise. Since motor actions are defined by the tasks they accomplish, the authors argue that mathematical tools for identifying muscle modules should account for task information from the outset.

Methods: The DsNM3F Algorithm

The researchers introduced DsNM3F (Discriminative sample-based Non-negative Matrix tri-Factorization), an algorithm that builds upon the "space-by-time" model. This model decomposes EMG data into three core components:

  1. Spatial Modules: Muscle groupings that activate together.
  2. Temporal Modules: Specific activation patterns over time.
  3. Activation Coefficients: The weighting of each module combination for a specific trial.
The DsNM3F framework decomposes EMG data into spatial and temporal modules while simultaneously optimizing for data reconstruction and task discriminability using a dual-objective cost function.

The innovation of DsNM3F lies in its dual-objective cost function. While standard methods only minimize reconstruction error, DsNM3F incorporates a Linear Discriminant Analysis (LDA) objective. This penalizes the model if activation coefficients for the same task are too scattered or if coefficients for different tasks are too similar. This "task-awareness" ensures that the extracted modules are functionally relevant.

Experimental Validation

The algorithm was validated using both simulated and experimental datasets:

  • Simulated EMG: A synthetic dataset with 40% added noise to test the recovery of ground-truth modules.
  • Experimental EMG: Data from 9 upper limb muscles in six healthy subjects performing 16 distinct arm-pointing tasks (8 directions at 2 speeds).

Key Results and Insights

The study revealed that adding task discrimination does not drastically alter the shape of spatial or temporal modules. Instead, it primarily refines the activation coefficients, suggesting that while the "building blocks" of movement are consistent, their combination is highly task-specific.

Quantitative comparison of DsNM3F against original data and the non-discriminative sNM3F algorithm, highlighting the trade-off between Variance Accounted For (VAF) and Discrimination (DEC).

Key performance metrics include:

  • Perfect Categorization: DsNM3F achieved near-perfect (1.0) task decoding on training sets, significantly outperforming the non-discriminative sNM3F (~0.75).
  • Minimal Accuracy Loss: The Variance Accounted For (VAF) dropped only slightly (e.g., from 1.0 to 0.82 in the example shown), indicating that task-awareness does not compromise reconstruction quality.
  • Generalization: The modules extracted by DsNM3F generalized well to unseen data, proving they represent physically meaningful structures rather than mathematical artifacts.

Conclusion and Future Directions

DsNM3F demonstrates that incorporating task-discriminative objectives is highly effective for describing the task modulation of module recruitment. While currently limited to discrete tasks, future work aims to extend this framework to continuous parameters like 3D coordinates. This approach holds significant promise for Brain-Machine Interfaces (BMIs) and rehabilitation robotics, where mapping muscle activations to intended tasks is critical for intuitive control and targeted therapy.

Summary updated

Abstract

Movement generation has been hypothesized to rely on a modular organization of muscle activity. Crucial to this hypothesis is the ability to perform reliably a variety of motor tasks by recruiting a limited set of modules and combining them in a task-dependent manner. Thus far, existing algorithms that extract putative modules of muscle activations, such as Non-negative Matrix Factorization (NMF), identify modular decompositions that maximize the reconstruction of the recorded EMG data. Typically, the functional role of the decompositions, i.e., task accomplishment, is only assessed a posteriori. However, as motor actions are defined in task space, we suggest that motor modules should be computed in task space too. In this study, we propose a new module extraction algorithm, named DsNM3F, that uses task information during the module identification process. DsNM3F extends our previous space-by-time decomposition method (the so-called sNM3F algorithm, which could assess task performance only after having computed modules) to identify modules gauging between two complementary objectives: reconstruction of the original data and reliable discrimination of the performed tasks. We show that DsNM3F recovers the task dependence of module activations more accurately than sNM3F. We also apply it to electromyographic signals recorded during performance of a variety of arm pointing tasks and identify spatial and temporal modules of muscle activity that are highly consistent with previous studies. DsNM3F achieves perfect task categorization without significant loss in data approximation when task information is available and generalizes as well as sNM3F when applied to new data. These findings suggest that the space-by-time decomposition of muscle activity finds robust task-discriminating modular representations of muscle activity and that the insertion of task discrimination objectives is useful for describing the task modulation of module recruitment.
BibTeX
@inproceedings{delis2015task,
  author = {Ioannis Delis and Stefano Panzeri and Thierry Pozzo and Bastien Berret},
  title = {{Task-discriminative space-by-time factorization of muscle activity}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2015},
  doi = {10.3389/fnhum.2015.00399},
}
Figure