Neurodevelopmental Outcome in Very Low Birth Weight Preterm Infants: An Exploratory Multivariable Analysis Including Sonographic Brain Volume Trajectories—Data from the NeoNEVS Project

Background: Extremely and very preterm infants are at high risk for adverse neurodevelopmental outcomes. Early prediction remains challenging when relying on static clinical markers or single time-point neuroimaging. Serial cranial ultrasound (CUS) enables repeated bedside assessment of cerebral growth and may provide longitudinal trajectory biomarkers integrable with routine clinical data. Methods: In this retrospective two-center cohort study, 89 preterm infants (<32 weeks’ gestation and/or <1500 g birth weight) were assessed using the Bayley-III at 24 months corrected age. Brain volume trajectory features were derived from serial CUS using a standardized ellipsoid model. A three-level analytical framework was applied as follows: univariate regression (62 models, Bonferroni and Benjamini–Hochberg correction), multivariate SVM classification with five-fold GroupKFold cross-validation, ensuring patient-level data separation and feature importance analysis with interaction characterization using stratified Spearman correlation and two-dimensional partial dependence plots. Results: Multivariate classification yielded modest but above-chance performance (balanced accuracy 0.277–0.463, Cohen’s κ 0.042–0.152). Respiratory morbidity duration—mechanical ventilation and BPD severity—were the most robustly associated univariate predictors, surviving Bonferroni correction. Brain volume trajectory features showed no significant univariate associations but contributed conditionally within the multivariate framework as follows: the interaction between brain volume slope and trajectory linearity was the strongest for cognitive outcome (Δr = 0.47), and postnatal growth restriction showed amplified adverse effects at lower birth weight for motor outcome (Δr = 0.47). Conclusions: This study demonstrates the value of ML methods as structured analytical tools for characterizing predictor–outcome relationships in preterm neurodevelopment; respiratory morbidity and brain volume trajectory features emerged as the most informative predictor classes. Prospective multicenter validation is required before clinical translation.