Improving Noise Robustness in EEG Source Localization with Sparse Bayesian Learning
Subhashini K. J., Ramkumar S · 2025
Extracting the neural correlates for observed scalp signals is perhaps the most intricate yet crucial stage of the process of EEG source localization. This essentially creates a neuroimaging component. However, as sLORETA and MNE methodologies are clinically tailored, they suffer from lack of relevant spatial resolution and highly constrained noise. An SBL framework aimed at noise robustness and the precision of localization relative to EEG source localization benchmarks is presented in this paper. This approach combines enhancement refinements of the EM algorithm with Bayesian hierarchical models that increase prior sparsity relative to the estimate of source amplitude which improves estimation. The SBL framework was compared to MNE and sLORETA using the BCI Competition IV EEG dataset scaled to UCS-II. The analysis demonstrated substantial improvement using SBL over the baseline methods of 5.12 mm LE and. 089 RMSE. Concomitantly, spatial assessment showed the frames achieved under SBL framework exhibit sharper activation contours beyond noise detectable contouring. The non-CNF from whence the activations originated enhanced precision while supremely constraining due to enforcing sparse priors robustly. Because of its accuracy and low sensitivity to imperfections, SBL framework's effectiveness in localization makes it particularly valuable for real-time neuroimaging applications.