Statistical selection of multiple-input multiple-output nonlinear dynamic models of spike train transformation

Area of Science:

• Computational neuroscience
• Neuroprosthetics
• Statistical modeling

Background:

• Nonlinear dynamic models, specifically Volterra kernel-based models, are used to describe spike train transformations in neural systems.
• Previous formulations exist for multiple-input multiple-output (MIMO) models in hippocampal-cortical prostheses.

Purpose of the Study:

• To describe statistical methods for selecting significant inputs (self-terms) and interactions (cross-terms) in Volterra kernel-based models.
• To develop reduced kernel models that are computationally efficient yet accurate for neural data.

Main Methods:

• A forward stepwise model selection technique was employed to determine model structure by progressively adding terms.
• The Wald test was utilized to prune model coefficients, reducing model complexity.
• The developed methods were applied to analyze functional interactions between neurons during behavior.

Main Results:

• Reduced kernel models with significantly fewer coefficients were developed.
• These reduced models demonstrated good fits to novel neural data.
• The selected terms effectively captured essential input self-terms and cross-terms.

Conclusions:

• The proposed statistical methods enable the creation of parsimonious Volterra kernel models for neural data.
• These reduced models offer a viable approach for analyzing functional neural interactions in neuroprosthetic applications.