Bayesian estimation of neural activity for non stationary sources using time frequency based priors

Electroencephalographic (EEG) recordings contain dynamic information inherent to its complex behavior, therefore, the accurate estimation of neural activity is highly dependent on the inclusion of such information in the inverse problem solution. The present work presents a way to obtain constraints for the Bayesian inverse problem solution, through a Variational Bayes Approach, using information contained in the space-time-frequency, more specifically under the Automatic Relevance Determination (ARD) framework. The time-frequency representation of the EEG allows to extract information that could be hidden in the nonstationarities and noise that are usually present in EEG data. The performance of the proposed method is evaluated using simulated EEG data under several SNRs in terms of spatial accuracy, temporal accuracy and mean squared error. Obtained results show that the proposed approach improves the spatial accuracy of the inversion under low SNR-data i.e., it is more robust to noise. Nevertheless, it does not show any improvement in the temporal accuracy of the estimations. Furthermore, the mean squared error does not show any significant result to assess the performance of the inversions.