117. fast and robust sparse bayesian learning for eeg source imaging

Department: Electrical & Computer Engineering
Faculty Advisor(s): Kenneth Kreutz-Delgado | Kenneth Kreutz-Delgado

Primary Student
Name: Alejandro Ojeda
Email: alojeda@ucsd.edu
Phone: 619-581-3548
Grad Year: 2019

We propose a new Sparse Bayesian Learning (SBL) algorithm that can deliver fast and robust solutions to the EEG source imaging (ESI) problem in the presence of highly noisy measurements. Images obtained by this algorithm result in brain activations that are spatially smooth and sparse. SBL is a probabilistic framework in which sparse solutions are obtained by optimizing the model evidence. In ESI, the evidence generally depends on two key unknown components: 1) the measurements noise covariance and 2) the sparsity profile of the solution. Popular implementations of the SBL framework often display suboptimal convergence properties, incurring an elevated computational cost and performance degradation due to poor data quality. In this paper, we address these shortcomings from a model selection perspective by maximizing the Bayes factor (evidence ratio) between a full and a sparse model in two steps. At every step, we use a convex approximation to the evidence as the surrogate cost function. First, we optimize the full model, yielding as a byproduct an estimate of the measurement noise covariance. Second, we maximize the evidence further by optimally removing irrelevant sources. We validated our method on both simulated and real EEG data. Our simulations show that the method is robust to measurement noise while outperforming in speed a number of popular EEG inverse solvers. On real data, the images obtained are in agreement with the experimental literature, being able to compactly localize the sources of error-related negative/positive potential to the anterior/posterior cingulate cortex.

Industry Application Area(s)
Life Sciences/Medical Devices & Instruments | Software, Analytics | Neuroimaging

« Back to Posters or Search Results