Title: An adaptive multilayer basis approach for task fMRI data
Authors: Michelle Miranda - University of Victoria (Canada) [presenting]
Jeffrey Morris - University of Pennsylvania (United States)
Abstract: A novel model is proposed to analyze task fMRI data that learns information in an adaptive way, carefully considering the complex structure of the brain data. The approach provides a sparse spatial representation of the brain while yielding full Bayesian inference at the voxel and ROI level with incredible computational speed. In addition, the proposed model allows for free full Bayesian inference on the residual connectivity, which can help scientist gain insights of the underlying brain function. We incorporate biological information from the brain's Regions of Interest (ROIs), accounting for both local correlation in each ROI and distant correlation between ROIs. Time dependencies in the BOLD time series are also considered by projecting the time course into a wavelet space. We then model the final set of bases by assuming a long memory process that accounts for differences in each wavelet decomposition level. We present a simulation study that shows increased power to detect activation when using the proposed composite-hybrid approach. We apply our method to the Working Memory task data from the Human Connectome Project.