Improving EEG Source Localization Through Spatio-Temporal Sparse Bayesian Learning

Sparse Bayesian Learning (SBL) approaches to the EEG inverse problem such as Champagne have been shown to outperform traditional ℓ
1
-norm based methods in terms of reconstructing sparse source configurations. Current approaches are however sensitive to strong noise contributions and assume independent samples, whereas neurophysiological time series are strongly auto-correlated. Here we present extensions, backed by compressive sensing theory, to the Champagne algorithm that improve the reconstruction performance in low-SNR settings as well as in the presence of correlated measurements. Our numerical simulations using a realistic EEG forward model confirm the efficacy of our approaches.