Variational Bayesian inversion for hierarchical unsupervised generative embedding (HUGE).

Area of Science:

• Neuroimaging
• Computational Psychiatry
• Machine Learning

Background:

• Hierarchical generative models offer a unified approach for effective connectivity inference and subgroup discovery in fMRI data.
• The existing hierarchical unsupervised generative embedding (HUGE) method, using dynamic causal modelling (DCM) and Markov chain Monte Carlo (MCMC) sampling, faces computational limitations.

Purpose of the Study:

• To develop an efficient variational Bayesian (VB) inversion scheme for the HUGE model.
• To accelerate the computational speed of HUGE while maintaining analytical accuracy.
• To enable practical applications of HUGE in clinical neuromodeling and computational psychiatry.

Main Methods:

• Derivation of a variational Bayesian (VB) inference scheme for the HUGE model.
• Validation using synthetic fMRI datasets with known ground truth.
• Evaluation on an empirical fMRI dataset from stroke patients and healthy controls.

Main Results:

• The VB scheme for HUGE demonstrated a two-orders-of-magnitude speed-up in model inversion compared to MCMC.
• The VB approach maintained a similar level of accuracy to MCMC.
• The VB implementation of HUGE is sufficiently fast for multi-start procedures, aiding whole-group analyses.

Conclusions:

• The VB inversion scheme provides a computationally efficient and accurate method for HUGE.
• This advancement facilitates the unsupervised detection of connectivity-defined subgroups in heterogeneous populations.
• HUGE, with VB inference, offers a practical solution for clinical neuromodeling and computational psychiatry.