Model selection for the extraction of movement primitives.

Area of Science:

• Motor Control
• Computational Neuroscience
• Biophysics

Background:

• Blind source separation (BSS) methods are crucial for extracting movement primitives from electromyography (EMG) and kinematic data in motor control research.
• Existing BSS model selection (e.g., PCA, ICA) often relies on heuristic approaches, potentially biasing results.
• There is a need for objective criteria to guide the selection of BSS model type, number of primitives, and smoothness priors.

Purpose of the Study:

• To develop and validate an objective criterion for selecting appropriate blind source separation models in motor control.
• To determine the optimal number of movement primitives and temporal smoothness constraints for BSS models.
• To provide a data-driven approach for model selection, moving beyond heuristic methods.

Main Methods:

• Re-formulated BSS models as Bayesian generative models.
• Employed a Laplace approximation to the posterior distribution of model parameters.
• Validated the proposed criterion against ground truth data and traditional model selection criteria (BIC, AIC).

Main Results:

• The proposed objective criterion demonstrated performance comparable to or exceeding traditional methods like BIC and AIC on ground truth data.
• Analysis of human gait data revealed that an anechoic mixture model with temporal smoothness on sources provided the best fit.
• The criterion effectively guided the selection of model type, number of primitives, and smoothness priors.

Conclusions:

• The developed Bayesian criterion offers an objective and effective method for selecting blind source separation models in motor control research.
• This approach enhances the reliability and interpretability of extracted movement primitives from EMG and kinematic data.
• An anechoic mixture model with temporal smoothness is suggested as a suitable model for analyzing human gait patterns.