Removal of the ballistocardiographic artifact from EEG–fMRI data: a canonical correlation approach