Comparing variational Bayes with Markov chain Monte Carlo for Bayesian computation in neuroimaging

Area of Science:

• Neuroscience
• Computational Statistics
• Brain Imaging

Background:

• Brain imaging studies generate complex, large datasets requiring sophisticated statistical models.
• The neuroelectromagnetic inverse problem in electroencephalography (EEG) involves estimating brain activity from scalp sensor data.
• Underdetermined dynamic linear models are often used, posing challenges for parameter estimation.

Purpose of the Study:

• To explore Bayesian computation methods for complex brain imaging data.
• To specifically address the neuroelectromagnetic inverse problem in EEG using variational approximations.
• To compare the accuracy and computational efficiency of variational methods against Markov chain Monte Carlo (MCMC).

Main Methods:

• Developed variational approximations for fitting hierarchical models.
• Derived approximations for a simple distributed source model and a complex spatiotemporal mixture model.
• Compared variational approximations with MCMC using synthetic and real EEG data.

Main Results:

• Variational methods demonstrated significant computational advantages for fitting EEG models.
• The accuracy of the derived variational approximations was clarified.
• The study analyzed an EEG dataset related to face perception to validate the methods.

Conclusions:

• Variational approximations offer an efficient and accurate approach for Bayesian computation in EEG source localization.
• These methods provide a valuable alternative to MCMC, particularly for large-scale neuroimaging datasets.
• The findings encourage wider adoption of variational methods in statistical neuroimaging research.