A hierarchical model for integrating unsupervised generative embedding and empirical Bayes

Area of Science:

• Neuroimaging analysis
• Computational neuroscience
• Statistical modeling

Background:

• Dynamic causal models (DCMs) infer effective connectivity from neuroimaging data.
• Generative embedding uses DCM parameters for patient classification and subgroup discovery.

Purpose of the Study:

• To present a novel hierarchical framework unifying individual connectivity inference with population structure discovery.
• To enable empirical Bayesian estimates of subject connectivity using subgroup priors.
• To introduce a Markov chain Monte Carlo method for model inversion.

Main Methods:

• Combined DCMs with finite mixture models into a single hierarchical model.
• Developed a Markov chain Monte Carlo sampling method for hierarchical generative model inversion.
• Applied the framework to simulated data and an empirical fMRI dataset (healthy controls and schizophrenia patients).

Main Results:

• Demonstrated face validity using simulated and empirical fMRI data.
• Achieved superior model evidence compared to conventional non-hierarchical DCM inversions on empirical fMRI data.

Conclusions:

• Presented a novel unified framework for joint inference of effective connectivity and population structure.
• The mixture model approach enables discovery of connectivity-defined clusters within a population.