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