Maximum likelihood estimation applied to multiepoch MEG/EEG analysis

A maximum likelihood based algorithm for reducing the effects of spatially colored noise in evoked response MEG and EEG experiments is presented. The signal of interest is modeled as the low rank mean, while the noise is modeled as a Kronecker product of spatial and temporal covariance matrices. The temporal covariance is assumed known, while the spatial covariance is estimated as part of the algorithm. In contrast to prestimulus based whitening followed by principal component analysis, our algorithm does not require signal-free data for noise whitening and thus is more effective with non-stationary noise and produces better quality whitening for a given data record length. The efficacy of this approach is demonstrated using simulated and real MEG data.; Next, a study in which we characterize MEG cortical response to coherent vs. incoherent motion is presented. It was found that coherent motion of the object induces not only an early sensory response around 180 ms relative to the stimulus onset but also a late field in the 250--500 ms range that has not been observed previously in similar random dot kinematogram experiments. The late field could not be resolved without signal processing using the maximum likelihood algorithm. The late activity localized to parietal areas. This is what would be expected. We believe that the late field corresponds to higher order processing related to the recognition of the moving object against the background.; Finally, a maximum likelihood based dipole fitting algorithm is presented. It is suitable for dipole fitting of evoked response MEG data in the presence of spatially colored noise. The method exploits the temporal multiepoch structure of the evoked response data to estimate the spatial noise covariance matrix from the section of data being fit, eliminating the stationarity assumption implicit in prestimulus based whitening approaches. The preliminary results of the application of this algorithm to the simulated data show its robustness to relatively high levels of noise. The bootstrap technique was used to assess the effectiveness of the algorithm on real data.