Practical limits on muscle synergy identification by non-negative matrix factorization in systems with mechanical...

Area of Science:

• Neuroscience
• Motor Control
• Computational Biology

Background:

• Statistical decomposition methods like Non-negative Matrix Factorization (NMF) are used to analyze behavioral motor programs.
• The relationship between factors identified by NMF and underlying neural structures remains unclear.
• NMF factors are typically extracted from a limited, low-dimensional command space.

Purpose of the Study:

• To investigate how Non-negative Matrix Factorization (NMF) factors relate to actual neural structures in motor control.
• To assess the impact of limited sampling within the control space on NMF's ability to recover underlying control modules.
• To evaluate NMF's performance in extracting control modules from simulated muscle activation patterns under different constraints.

Main Methods:

• Applied Non-negative Matrix Factorization (NMF) to muscle activation patterns synthesized from low-dimensional, synergy-like control modules.
• Simulated two paradigms: task-constrained control space and feedback activation from proprioceptive signals.
• Assessed factor accuracy based on the volume of control space sampled and the number of internal degrees of freedom in a mechanical model.

Main Results:

• In the task-constrained paradigm, NMF factor accuracy significantly degraded with reduced sampling of the control space (e.g., 50% sampling caused substantial degradation).
• In the feedback paradigm, NMF failed to extract more than four control modules, even in a model with seven degrees of freedom.
• Physical constraints limiting access to the low-dimensional control space led to distorted recovered factors.

Conclusions:

• The accuracy of NMF in recovering control modules is highly sensitive to the extent of control space sampling.
• Physical constraints can substantially distort low-dimensional controllers, leading to inaccurate dimensionality and composition of extracted factors.
• NMF-derived factors may not accurately represent the true underlying neural control structures when data sampling is limited or constrained.