Mrinet: a Morphometric Relationship Inference Network for the Prediction of Neonatal Neurodevelopmental Outcomes at 2 Years

Preterm infants are accompanied with marked brain developmental delays, making them vulnerable to neurodevelopmental disorders, such as atypical linguistic and motor performances. Accurate prediction of neurodevelopmental outcomes at 2 years of age using brain metrics in neonates is of great importance for identifying potential developmental risks associated with conditions at birth. The study introduces a new Morphometric Relationship Inference Network (MRINet) to forecast linguistic and motor performances in two-year-olds. MRINet uses a Transformer encoder alongside a relational inference component, effectively modeling both distant and proximate multimodal dependencies within cortical regions. Our model outperformed others in predicting linguistic and motor performance of infants at 18-24 months by using their brain features at birth, with mean absolute errors of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.08 \pm 2.12$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.74 \pm 1.24$</tex>, respectively, and Pearson's correlations over 0.72 between the predicted scores and original assessments. These findings highlight the effectiveness of the MRINet, offering promising clinical applications for identifying neurodevelopmental risks early in the perinatal phase.