Swin fMRI Transformer Predicts Early Neurodevelopmental Outcomes from Neonatal fMRI

Early intervention potential: The SwiFT model successfully predicts Bayley-III developmental scores (cognitive, motor, and language outcomes) from neonatal brain scans, which could enable clinicians to identify developmental delays earlier and provide timely interventions during a critical period of brain development.

Addressing data scarcity through transfer learning: The researchers tackled a common responsible AI challenge—limited neonatal data—by pretraining their model on large adult fMRI datasets, demonstrating how transfer learning can make AI systems more practical and reliable when working with sensitive pediatric populations where data collection is inherently limited.

Interpretability for clinical trust: The team used Integrated Gradients with Smoothgrad sQuare (IG-SQ) to identify which specific brain regions the model relied on for predictions, moving beyond "black box" AI by showing which neural patterns actually correlate with developmental outcomes—a crucial step for clinicians to understand and trust AI recommendations.

Multi-outcome prediction approach: Rather than predicting a single outcome, SwiFT uses both single-label and multi-label strategies to predict cognitive, motor, and language development simultaneously, reflecting the interconnected nature of early development and providing more comprehensive clinical insights than simpler models.

Superior performance with architectural innovation: The Transformer-based architecture significantly outperformed baseline models by processing spatiotemporal brain data end-to-end, suggesting that more sophisticated AI approaches can meaningfully improve prediction accuracy in high-stakes medical applications involving vulnerable populations.