Canonical polyadic and block term decompositions to fuse EEG, phenotypic scores, and structural MRI of children with early-onset epilepsy

Abstract

We investigated two popular tensor decomposition models, canonical polyadic decomposition (CPD) and block term decomposition (BTD), to test their ability to fuse datasets from three different modalities related to neuroscience. We fused electroencephalogram (EEG) spectral power, regional brain volume from magnetic resonance imaging (MRI) and phenotypic scores from 29 preschool children aged <; 5 y.o. who have a diagnosis of epilepsy. We used CPD and BTD in a coupled matrix-matrix-tensor factorisation setting to find shared components across data modalities. In addition, we imposed a hard constraint on the model to extract factors directly interpretable in terms of childhood development. We evaluated the model performance to extract components in agreement with prior clinical knowledge. We found that both models revealed similar patterns of relationships between regional brain volumes and developmental scores following prior clinical knowledge but BTD was slightly more sensitive than CPD.