Multiple constrained minimum-variance beamformers for reconstruction of induced and evoked neural activity

Abstract

Event Abstract Back to Event Multiple constrained minimum-variance beamformers for reconstruction of induced and evoked neural activity Alexander Moiseev1* 1 Down Syndrome Research Foundation, Canada Linearly constrained minimal variance beamformers [1, 2] are widely used for brain source reconstruction. These adaptive spatial filters are highly effective for weakly correlated brain activity. However, when correlations become significant, filter performance drastically degrades [3]. One solution to this problem is to use a multiple constrained minimal variance (MCMV) beamformer [4] which requires precise a priori information about the sources. This information is usually not available. For this reason MCMV-based approaches that relax this restriction were developed [5, 6]. Yet questions about the application of the MCMV filter to find multiple unknown correlated sources and whether unbiased estimates of source locations can be obtained remain unanswered. To address these issues, we derived MCMV-based source localizers that can be applied to both induced and evoked activity (multi-source activity index, multi-source pseudo-Z) and those targeting only the evoked component of the field (full and reduced multi-source event-related beamformers). We show that for arbitrary noise covariance all these localizers remain bounded at singular points and yield simultaneous unbiased estimates of all source positions and orientations. We also derived a search algorithm that starts with a conventional single source beamformer then iteratively finds new sources and adds them to the MCMV forward solution. While this algorithm does not guarantee convergence to an exact solution, it provides a practical alternative to a brute force search which is unfeasible for realistic numbers of interfering sources. The proposed multi-source localizers were tested on simulated data with multiple correlated dipoles and on real MEG data collected in an auditory steady state response experiment. We showed that where conventional beamformers were unable to detect coherent sources, the MCMV-based ones could find most or all of them, depending on the SNR.