Identifying Subgroups and Novel Brain-Behavior Associations in Adolescents Who were Born Very Preterm Using Bi-clustering Approaches

Here we apply Bi-clustered Independent Component Analysis (B-ICA), a data driven two-dimensional clustering algorithm, on three-dimensional T1-weighted magnetic resonance images of 126 adolescents born very preterm (VPT; < 33 weeks of gestation; mean age=15.29 years; standard deviation=0.58). Two components of interest covering regions of insula and hippocampus showing persisting structural alterations following very preterm birth were chosen, based on previous literature. Corresponding component loadings were sorted using the positive or negative method as suggested by the N-way bi-clustering (n-BIC) algorithm. The association between component loadings and participants’ total scores on the Rutter Parents Scale (RPS), a scale measuring emotional and conduct difficulties, was studied in three data- driven subgroups. Loadings of subgroup Sinter (n=32), weighted on both insula and hippocampus imaging components, correlated with RPS Total scores (r=-0.458; p=0.008). Subgroup S1 (n=29), heavily weighted on the insula component, did not show any association between regional brain volume and behavioral outcome. However, for the same S1 subgroup, loadings on the hippocampal component correlated with RPS Total scores (r =-0.61; p=3.43e-4). No significant associations between regional brain volume and behavioral outcome were observed for subgroup S2 (n=34). These results suggest that distinct subgroups of VPT adolescents with similarly increased emotional and conduct problems may have differential underlying brain mechanisms implicating different parts of large-scale emotional and cognitive systems. This study enables interpretation of the neurodevelopmental sequelae in VPT by characterizing homogenous subgroups using data driven multivariate methods. Our results could guide in determining subgroups of VPT born adolescents at the risk of developing behavioral issues as they grow-up.