DYNLO: Enhancing Non-linear Regularized State Observer Brain Mapping Technique by Parameter Estimation with Extended Kalman Filter

Abstract

The underlying activity in the brain can be estimated using methods based on discrete physiological models of the neural activity. These models involve parameters for weighting the estimated source activity of previous samples, however, those parameters are subject- and task-dependent. This paper introduces a dynamical non-linear regularized observer (DYNLO), through the implementation of an Extended Kalman Filter (EKF) for estimating the model parameters of the dynamical source activity over the neural activity reconstruction performed by a non-linear regularized observer (NLO). The proposed methodology has been evaluated on real EEG signals using a realistic head model. The results have been compared with least squares (LS) for model parameter estimation with NLO and the multiple sparse prior (MSP) algorithm for source estimation. The correlation coefficient and relative error between the original EEG and the estimated EEG from the source reconstruction were inspected and the results show an improvement of the solution in terms of the aforementioned measurements and a reduction of the computational time.