Abstract Dramatic alterations in brain morphology, such as cortical thickness and sulcal folding, occur during the 3rd trimester of gestation which overlaps with the period of premature births. Here, we investigated the ability of the graph convolutional network (GCN) to predict brain age for preterm neonates by accounting for morphometrics measured on the cortical surface and the surface mesh topology as a sparse graph. Our findings demonstrate that GCN-based age prediction of preterm neonates (n=170; mean absolute error [MAE]: 1.06 weeks) outperformed conventional machine learning algorithms and deep learning methods that did not use topological information. We further evaluated how predicted brain age (PBA) emerges as a biologically meaningful index that characterizes the current status of brain development at the time of imaging. We hypothesized that the relative brain age (RBA; PBA minus chronological age) at scan reflects a combination of perinatal clinical factors, including preterm birth, birthweight, perinatal brain injuries, exposure to postnatal steroids, etc. We also hypothesized that RBA of neonatal scans may be associated with brain functional development in the future. To validate these hypotheses, we used general linear models. Furthermore, we established structural equation models (SEM) to determine the structural relationship between preterm birth (as a latent variable of birthweight and birth age), perinatal injuries (as a latent variable of three leading brain injuries), postnatal factors (as a latent variable of six clinical conditions), RBA at scan, and neurodevelopmental scores at 30 months. Our results suggest that low birthweight, chronic lung disease, and exposure to postnatal steroids impair cortical growth, as low RBA was significantly associated with these risks. Furthermore, RBA was associated with cognitive and language scores at 30 months. SEM analysis indicated that RBA mediated the influences of preterm birth and postnatal clinical factors, but not perinatal brain injuries, toward brain functional development at 30 months. The left middle cingulate cortex showed the most accurate prediction of brain age (MAE: 1.19 weeks), followed by left posterior and right middle cingulate cortices (1.21 weeks). These cingulate regions presented faster growth than others. RBAs of several frontal cortices significantly correlated with cognitive abilities at 30 months of age (n=50). Whereas, RBA of left Broca’s area, which is important for language production and comprehension, was associated with language functional scores. Overall, our results demonstrate the potential of the GCN in both predicting brain age and localizing regional growth that relates to postnatal factors and future neurodevelopmental outcome.

Brain age predicted using graph convolutional neural network explains developmental trajectory in preterm neonates