Bayesian model averaging in EEG/MEG imaging

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• The electroencephalography/magnetoencephalography (EEG/MEG) inverse problem (IP) seeks to localize brain activity from scalp recordings.
• Existing inverse methods face challenges like model uncertainty, ghost sources, and underestimation of deep brain activity.

Purpose of the Study:

• To develop a novel Bayesian framework for the EEG/MEG inverse problem using Bayesian Model Averaging (BMA).
• To address model uncertainty and improve the accuracy of Primary Current Density (PCD) estimation.
• To mitigate issues such as ghost sources and underestimation of deep brain activity in linear inverse solutions (LIS).

Main Methods:

• Formulation of the EEG/MEG inverse problem using Bayesian Theory.
• Introduction of a third level of inference: Bayesian Model Averaging (BMA).
• Application of BMA to EEG IP in the frequency domain with varying anatomical constraints.
• Validation using simulated and real experimental data.

Main Results:

• The BMA approach effectively handles model uncertainty in the EEG/MEG inverse problem.
• Demonstrated reduction in ghost sources and improved estimation of deep brain activity compared to traditional LIS.
• BMA solutions showed competitive performance against LORETA and cLORETA.

Conclusions:

• Bayesian Model Averaging offers a robust framework for solving the EEG/MEG inverse problem.
• This methodology enhances the reliability and accuracy of brain source localization.
• BMA provides a promising alternative for analyzing EEG/MEG data, particularly in complex scenarios.