Time-varying dynamic Bayesian network learning for an fMRI study of emotion processing.

Area of Science:

• Computational neuroscience
• Machine learning
• Network analysis

Background:

• Dynamic Bayesian Networks (DBNs) are crucial for modeling time-varying systems.
• Learning DBNs, especially with complex, time-varying structures, remains a significant challenge.
• Existing methods often struggle with high-dimensional data and dynamic network changes.

Purpose of the Study:

• To present a novel, scalable method for learning time-varying dynamic Bayesian networks.
• To address the challenges of high dimensionality, time-varying structures, and multi-subject data in DBN learning.
• To improve estimation accuracy and computational efficiency compared to existing approaches.

Main Methods:

• Decomposition of the DBN learning problem into a sequence of regression inference problems.
• Application of Markov neighborhood regression for each inference problem.
• Validation through extensive numerical experiments and application to fMRI data.

Main Results:

• The proposed method demonstrates scalability with data dimensionality.
• It effectively accommodates time-varying network structures and handles multi-subject data.
• Numerical experiments confirm superior performance in estimation accuracy and computational efficiency over existing methods.

Conclusions:

• The novel method provides a robust and efficient approach for learning time-varying dynamic Bayesian networks.
• Application to fMRI data during an emotion-processing task highlights the subcortical-cerebellum's key role.
• This work advances the analysis of dynamic brain connectivity.