“Statistical Signal Processing of Magnetoencephalography Data”

by Dimitrios Pantazis

December 2006

Imaging approaches in MEG typically generate dynamic current density maps (CDMs) on the cortical surface, where the activation in each location is represented by a current dipole. The highly convoluted nature of the human cortex requires the use of many thousands of dipoles for an accurate representation of the cortical surface, and the resulting CDMs are difficult to interpret. As with fMRI images, objective assessment of CDMsrequires a principled approach to identifying regions of activation, which involves testing thousands of hypotheses (one per surface element) for statistically significant experimental effects. This raises the possibility of a large number of false positives simply as a result of multiple hypothesis testing. To overcome this problem, we develop statistical procedures that control the familywise error rate, i.e. the chance of one or more falsepositives under the null hypothesis of no activation. We use random field and permutation methods that consider the spatial dependence of the data, and make inferences based on the global maximum distribution of the image statistics. To satisfy random field assumptions, we must first smooth the CDMs. Since permutation methods are easily generalizable to multidimensional data, we expand them to control the familywiseerror rate over spatial-temporal-spectral maps. We compare and evaluate the methods on simulated and experimental data, and use them to detect hemispheric language dominance in epileptic patients and morphological differences in brain structures.

Since experimental designs often involve multiple factors and confounding effects, we present the use of general linear modeling theory to model MEG experimental data. We demonstrate the theory in a specific MEG experiment involving a visual spatial cueing paradigm. To extract contrast statistics of attention related alpha activity, we follow a massively univariate approach and model each subject, region of interest, andtime-frequency band separately with a novel ANCOVA (analysis of covariance) design, where power in rectangular time-frequency bands forms the observation variables, and baseline power forms the covariate. We also demonstrate the use of false discovery rate, a less conservative alternative to familywise error rate control, to do multiplicity adjustments to our statistics.

This report is not currently available in PDF format for downloading. Contact the Signal and Image Processing Institute for information on its availability.