Improving brain decoding through constrained and parametrized temporal smoothing

Decoding mental states from task-related fMRI data has recently been the focus of much research. Nevertheless, high levels of acquisition and physiological noise still makes inter-subject decoding a difficult and quite unstable process. Since all of the existing decoding approaches are applied on a volume-by-volume basis, it would be sensible to ensure that sudden signal changes reflect a true change of cognitive state rather than noise artefacts. Correction of the temporal signal can be achieved through temporal smoothing, which over the years has always been a debatable fMRI preprocessing step among the neuroscience community. In this paper, we present two methods for improving decoding accuracy by correcting the temporal dynamics of a number of functional regions, using parametrized temporal smoothing. We test our methods on a real fMRI dataset and we show that when temporal smoothing is applied separately in multiple scales and is both properly constrained and conditioned, it can remove sudden artefact-driven peaks and drops from the fMRI signal and thus improve the prediction accuracy of different tasks. Moreover, since our methods are performed independently from the decoding operations, they can be used in conjunction with any feature selection and classification algorithm.