Bayesian EEG source localization using a structured sparsity prior.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• Electroencephalography (EEG) is crucial for non-invasive brain activity monitoring.
• Accurate EEG source localization is challenging due to signal complexity and inverse problem nature.
• Existing methods may lack spatial coherence or robust temporal waveform recovery.

Purpose of the Study:

• To develop a novel hierarchical Bayesian model for spatially coherent focal EEG source localization.
• To recover temporal EEG waveforms accurately for potential clinical applications.
• To improve the robustness and recovery rate compared to existing regularization techniques.

Main Methods:

• Proposed a hierarchical Bayesian model incorporating a multivariate Bernoulli Laplacian structured sparsity prior.
• Implemented a partially collapsed Gibbs sampler for posterior sampling and joint estimation of brain activity and hyperparameters.
• Introduced two Metropolis-Hastings moves to accelerate sampler convergence: multi-dipole shifts and cross-chain proposals.

Main Results:

• The proposed algorithm demonstrated higher robustness and recovery rates on synthetic data compared to weighted ℓ21 mixed norm regularization.
• On real EEG data, the method achieved spatially coherent source localization comparable to state-of-the-art techniques (Multiple Sparse Prior, Champagne algorithm).
• The algorithm successfully estimated EEG waveforms with meaningful temporal peaks, valuable for characterizing activity spread.

Conclusions:

• The novel Bayesian approach offers improved performance in EEG source localization and waveform recovery.
• The method provides a robust and accurate tool for analyzing brain activity, with potential clinical utility.
• This work advances the field of EEG source imaging through a sophisticated Bayesian framework.