A Bayesian marginalization approach for improved EEG dipole localization

A Bayesian framework is utilized to derive a new EEG dipole localization method. Dipole orientation/amplitude parameters are integrated out of the full observation probability density function, yielding a proper marginal. The derived cost function essentially reduces to the condensed conditional maximum likelihood for a large number of electrodes, relative to the number of sources. In addition, the reduction of nuisance parameters by marginalization leads naturally to a model selection/source number detection scheme based on the concept of Bayesian evidence. Simulation results are included to show the superior estimation characteristics of the proposed method over maximum likelihood in low SNR, small spatial/temporal sample situations, and to provide a preliminary comparison of the proposed detection technique to AIC and MDL