Swin fMRI Transformer Predicts Early Neurodevelopmental Outcomes from Neonatal fMRI

TL;DR

Neonatal fMRI offers a window into early brain development, but predicting Bayley-III composite scores from neonatal rs-fMRI remains challenging due to high dimensionality and limited data. The authors introduce SwiFT, a 4D Swin Transformer that processes spatiotemporal fMRI data end-to-end and integrate group ICA features and cross-age pretraining to boost predictive accuracy. SwiFT outperforms ROI-based baselines in single- and multi-label predictions across cognitive, language, and motor domains, with ICA-based reductions further enhancing performance. IG-SQ interpretation links predictive patterns to neurobiological networks, supporting clinical relevance for early detection and intervention.

Abstract

Brain development in the first few months of human life is a critical phase characterized by rapid structural growth and functional organization. Accurately predicting developmental outcomes during this time is crucial for identifying delays and enabling timely interventions. This study introduces the SwiFT (Swin 4D fMRI Transformer) model, designed to predict Bayley-III composite scores using neonatal fMRI from the Developing Human Connectome Project (dHCP). To enhance predictive accuracy, we apply dimensionality reduction via group independent component analysis (ICA) and pretrain SwiFT on large adult fMRI datasets to address the challenges of limited neonatal data. Our analysis shows that SwiFT significantly outperforms baseline models in predicting cognitive, motor, and language outcomes, leveraging both single-label and multi-label prediction strategies. The model's attention-based architecture processes spatiotemporal data end-to-end, delivering superior predictive performance. Additionally, we use Integrated Gradients with Smoothgrad sQuare (IG-SQ) to interpret predictions, identifying neural spatial representations linked to early cognitive and behavioral development. These findings underscore the potential of Transformer models to advance neurodevelopmental research and clinical practice.

Figures (6)

• Figure 1: Distributions of composite scores based on binary classification. Assumptions made in johnson2014using largely hold true.
• Figure 2: Rationale of using large amounts of brain fMRI data across different datasets to make up for the little amount of data available for neonates. We expect that learned features from elderly, adult and child data will generalize to limited neonatal fMRI data and thus improve downstream performance for neonates.
• Figure 3: UML Flowchart depicting the workflow to extract features and connectivity maps by leveraging Group Independent Component Analysis ALFAROALMAGRO2018400, Miller2016 and GAL2022118920.
• Figure 4: Overview of the models' predictive power via 5-fold cross-validation for BSID-III regression ($\text{MAE}_{adj}$) and classification ($\text{ACC}_{bal}$). The strongest baseline was taken to represent baseline performance. Performance by transfer-learning and 100 ICs was omitted due to no statistically significant performance changes.
• Figure 5: A glimpse into network-level interpretations, focusing on IC 5 (left) possibly representing the Executive Control Network (ECN). We can see how this network adds to activation in the medial prefrontal cortex for prediction of cognitive delay (right).