Predicting neurodevelopmental outcomes from neonatal cortical microstructure: A conceptual replication study.

Area of Science:

• Neuroscience
• Developmental Science
• Machine Learning

Background:

• Machine learning and neuroimaging offer potential for predicting neurodevelopmental outcomes.
• Previous research suggested diffusion MRI fractional anisotropy (FA) predicts cognitive and language outcomes in neonates.

Purpose of the Study:

• To replicate and validate the predictive power of cortical microstructure (FA) for neurodevelopmental outcomes.
• To assess the generalizability of predictive models using larger, independent datasets.

Main Methods:

• Utilized the developing Human Connectome Project (dHCP) dataset with early MRI and Bayley Scales of Infant and Toddler Development, Third Edition (BSID-III) scores.
• Replicated a machine learning pipeline using fractional anisotropy (FA) from diffusion MRI.
• Validated the pipeline across multiple cohorts with varying gestational and post-menstrual ages.

Main Results:

• Replication attempts failed to achieve prediction accuracy beyond random levels.
• Negative results persisted even when expanding study settings and cohorts.
• Diffusion MRI-FA measures at birth were insufficient for reliable BSID-III score prediction in toddlerhood.

Conclusions:

• Cortical microstructure (DTI-FA) near birth may not be sufficient for predicting BSID-III scores in toddlerhood.
• Findings highlight challenges in replicating machine learning models in neurodevelopmental research.
• Emphasizes the need for robust validation and best practices in predictive tool development.