Modeling the hemodynamic response function using simultaneous EEG-fMRI data and convolutional sparse coding analysis with rank-1 constraints
Over the last few years, an increasing body of evidence points to the hemodynamic response function as an important confound of resting-state functional connectivity. Several studies in the literature proposed using blind deconvolution of resting-state fMRI data to retrieve the HRF, which can be subsequently used for hemodynamic deblurring. A basic hypothesis in these studies is that relevant information of the resting-state brain dynamics is condensed in discrete events resulting in large amplitude peaks in the BOLD signal. In this work, we showed that important information of resting-state activity, in addition to the larger amplitude peaks, is also concentrated in lower amplitude peaks. Moreover, due to the strong effect of physiological noise and head motion on the BOLD signal, which in many cases may not be completely removed after preprocessing, the neurophysiological origin of the large amplitude BOLD signal peaks is questionable. Hence, focusing on the large amplitude BOLD signal peaks may yield biased HRF estimates. To define discrete events of neuronal origins, we proposed using simultaneous EEG-fMRI along with convolutional sparse coding analysis. Our results suggested that events detected in the EEG are able to describe the slow oscillations of the BOLD signal and to obtain consistent HRF shapes across subjects under both task-based and resting-state conditions.