Predicting two-year neurodevelopmental outcomes in preterm infants using multimodal structural brain magnetic resonance imaging with local connectivity: A machine learning approach

Abstract

Abstract Determine brain structural networks in extremely preterm (EP; <28 weeks), very-to-late preterm (V-LP; ≥28 and < 37 weeks), and all preterm infants at term-equivalent age. Predict 2-year neurodevelopmental scores using multimodal predictors. Prospective cross-sectional study with MRI and diffusion MRI on 61 EP and 131 V-LP infants. Constructed a multimodal feature set through volumetric and structural network analysis. Linear and nonlinear machine learning models used for predicting Bayley Scales of Infant and Toddler Development, Third Edition (BSID-III) scores, assessing predictive power and feature importance. Prediction models, incorporating local connectivity features, demonstrated high predictive performance for cognitive scores in preterm (RMSE 13.352; variance explained 17%) and V-LP (RMSE 11.205; variance explained 17%) infants. For motor scores, models with local connectivity features had the highest predictive performance for EP (RMSE 11.363; variance explained 15%). A model with only local connectivity features showed high predictive performance for language scores in preterm infants (RMSE 11.792; variance explained 15%). BSID-III prediction performance and feature importance varied across preterm groups, emphasizing the efficacy of multimodal feature sets with local connectivity. Leveraging machine learning in this context enhances our understanding of microstructural alterations and their link to neurodevelopmental outcomes, facilitating risk stratification.