Comparison of SAM and MUSIC performance for unaveraged MEG

SAM (synthetic aperture magnetometry) and MUSIC (multiple signal classification) are methods for identifying dipolar sources represented in the covariance of MEG measurements. SAM operates by estimating source and noise power, as a function of position and current vector, from a full-rank covariance matrix [1]. MUSIC finds the locations for which a test dipole is orthogonal to the covariance noise subspace [2]. Thus, MUSIC requires the added nonlinear step of determining where to partition the covariance matrix into signal and noise subspaces. The presence of background brain activity implies that unaveraged MEG measurements have a full rank covariance matrix, with no noise subspace. Despite this, MUSIC has been successfully applied to unaveraged evoked activity, by prewhitening, using the background MEG activity [3]. In many applications, such as clinical MEG recordings of epileptic or focal slow wave activity, it is not always possible to identify MEG segments containing only background brain activity. Consequently, the prewhitening step of Sekihara cannot be easily applied. Nonetheless, it has been suggested that MUSIC can be applied to unaveraged MEG data, after judicious selection of the noise subspace partition. Our study compares the localization accuracy of SAM and MUSIC (for all feasible partitions) for unaveraged MEG. Our test cases include measurement of an electrolyte dipole phantom, for which dipole positions are accurately known, and clinical epilepsy data. The accuracy of both analyses was verified by the dipole phantom over a wide range of conditions. However, using interictal spike data, we were unable to find a rule for selection of the subspace partition rank at which all significant sources could be accurately localized simultaneously by MUSIC. Using the identical MEG data, SAM revealed multiple sources for which the localization uncertainty of each could be predicted from the value of the activation peaks. These uncertainties were small compared with the source localization scatter of MUSIC, corresponding to multiple noise subspace sizes. 2 Methods