Functional, structural, and phenotypic data fusion to predict developmental scores of pre-school children based on Canonical Polyadic Decomposition

Abstract

Recent technological advances enable the acquisition of diverse datasets that demand data-driven analysis. In this context, we seek to take advantage of diverse data modalities to explore the links between childhood development, structure and function of the brain. We deploy a data fusion model using coupled matrix-tensor decomposition of electroencephalography (EEG), structural magnetic resonance imaging (sMRI), and phenotypic score data to investigate how functional, structural, and phenotypic variables reflect development in young children with epilepsy. Our model is based on Canonical Polyadic Decomposition and optimised with grid search to predict developmental scores of pre-school children. The model is promising and able to show relationships between modalities that agree with clinical expectations. The score prediction yields a high similarity at the group level and potential to predict laborious and time-consuming developmental scores from routinely collected sMRI and/or EEG data, thus becoming a stepping-stone towards more efficient clinical assessment of brain development in young children.