Identification of fMRI Waveforms.

Abstract:
This study examines the utility of linear prediction modeling in assessing the likelihood that two fMRI time series represent the same underlying waveform. This knowledge can form the basis of novel approaches to studying brain function without making certain assumptions made by popular techniques, e.g., SPM. The fMRI data were generated from sets of BOLD images from a patient performing a unilateral fingertapping paradigm while being scanned by a GE echospeed Horizon1.5T Scanner using an echo pulse sequence(TR-2000 ms, TE-60 ms, 24 cm FOV, 5 mm slices, 14 slices). For each time series, auto-regressive linear prediction models were generated with an order indicated by a common regulizer such as the Akaike Information Criterion. The residual Q-Q plots and the percentage of explained variance were indices for the reliability of the models. Cross correlation statistical maps were used for comparison. The best fit models accounted for 55-75% of the variance and corresponded to voxels with high cross correlation values(>0.80). Adjacent voxels often differed in order or by coefficient values. Many current methods of fMRI analysis assume that the waveforms represent samples from a homogenous stationary linear process. Our results show that fMRI data are neither locally homogenous nor stationary but may be locally stochastic interacting linear dynamical processes(ARIMA results pending).