Methods for separating temporally overlapping sources of neuroelectric data

Abstract

The localization of intracranial sources of EEG or MEG signals can be misled by the combined effect of several sources, as illustrated by simulated MEG data in which two of the three dipolar sources have slightly out of phase activity and partly complementary scalp topographies. These data were analysed by three different source localization methods. Fitting a single source to each sequential topography worked perfectly when only one source was active; this could also account for as much as 95% of the spatial variance of topographies resulting from two overlapping sources, although the solution was then far from any source. A principal component analysis approach followed by an oblique rotation (fitting one source to the spatial aspect of each component) correctly localized two of the sources but severely mislocated the source that was never active alone. Spatio-temporal source modeling (simultaneously fitting a set of sources to all consecutive topographies) correctly localized all three sources, provided that the parameter optimization method could escape sub-optimal local minima of the error function. Temporally overlapping sources can thus be separated and correctly identified if the mathematical model is adequate and the optimization procedure is well adapted.

Recovering dipole sources from scalp-recorded event-related-potentials using component analysis: principal component analysis and independent component analysis

“…For example, under conditions of both temporal and spatial independence in the topography and activation of the generating sources, both methods predominantly resulted in components that could be adequately modeled with single-dipole models. There are some conditions for PCA, therefore, in which a single dipole (location and moment parameters) may be derived from single components for segmented EEG data (Maier et al, 1987;Achim et al, 1988; also cf. Mosher et al, 1992).…”

“…Thus, the eigenvector weights themselves could be used in equivalent current source analysis to model the dipoles that generated scalp-recorded EEG activity. It has been shown that the dipoles underlying the generation of scalp electrical activity may in some circumstances be recovered with PCA analysis (Maier et al, 1987;Achim et al, 1988;also cf. Mosher et al, 1992).…”

Functional and structural aspects of tinnitus-related enhancement and suppression of auditory cortex activity

“…PCA decomposes the original multivariate signal into component waveforms and associated weight vectors that determine how much each component contributes to the activity in each signal channel. By selecting the relevant components and excluding the irrelevant ones, the signal-to-noise ratio of the field distribution submitted to source analysis is improved (Achim et al, 1988;Diesch et al, 2004;Diesch and Flor, 2007;Lagerlund et al, 1997). The principal component representing the SSR components was identified on the basis of FFT-derived amplitude spectra.…”

Recursive MUSIC: A framework for EEG and MEG source localization

“…We assume simultaneous recordings at sensors for time instances. We can express the by spatio-temporal data matrix as (3) or (4) We refer to as the dipole "gain matrix" [12] that maps a dipole at into a set of measurements. The three columns of the gain matrix represent the possible forward fields that may be generated by the three orthogonal orientations of the th dipole at the sensor locations .…”

“…Early attempts at source localization were based on fitting the multiple-dipole model to a single time sample of the measurements across the E/MEG (EEG and/or MEG) array [5], [21], [31]. By noting that physiological models for the current sources typically assume that they are spatially fixed for the duration of a particular response, researchers were able to justify fitting the multipledipole model to a complete spatio-temporal data set [2], [3], [17], [18]. The spatio-temporal model can result in substantial improvements in localization accuracy; however, processing the entire data set leads to a large increase in the number of unknown parameters, since the time series for each source must now be estimated in addition to the dipole location and orientation.…”