Identification of sparse neural functional connectivity using penalized likelihood estimation and basis functions.

Area of Science:

• Computational Neuroscience
• Neural Engineering
• Statistical Modeling

Background:

• Identifying functional brain connectivity from spike train data is crucial but challenging.
• Model complexity and overfitting hinder interpretation in neural data analysis.
• Sparse representations are needed for efficient and interpretable functional connectivity models.

Purpose of the Study:

• To formulate a generalized functional additive model (GFAM) for sparse functional connectivity estimation.
• To develop penalized likelihood estimation methods for GFAMs.
• To address both global and local sparsity in neural network dynamics.

Main Methods:

• Developed a generalized functional additive model (GFAM) incorporating basis functions and link functions.
• Applied penalized likelihood estimation with group LASSO (global basis) and group bridge (local basis).
• Utilized quadratic approximation for model optimization and estimation.

Main Results:

• Both group-LASSO-Laguerre and group-bridge-B-spline GFAMs effectively captured global sparsity.
• The group-bridge-B-spline model accurately identified both global and local sparsities simultaneously.
• Sparse GFAMs demonstrated superior out-of-sample prediction performance compared to standard maximum likelihood models.

Conclusions:

• Sparse generalized functional additive models provide a powerful framework for analyzing neural connectivity.
• The proposed methods effectively reduce model complexity and improve interpretability.
• These models offer enhanced accuracy in predicting neural dynamics using spike train data.