Spatio-temporal reconstruction of early brain developmental trajectories via self-supervised learning.

Charting the normative and atypical trajectories of the brain's rapid early-life reorganization is crucial for understanding neurodevelopment. Although magnetic resonance imaging (MRI) enables high-resolution visualization of the structural and functional development, its application in longitudinal studies is limited by acquisition challenges, leading to severely incomplete follow-up data. To address this limitation, we propose STRIDER, a novel framework that integrates graph neural networks into the transformer architecture to reconstruct complete brain developmental trajectories from extremely sparse observations. We formulate this task as a multivariate time series imputation problem, enabling the estimation of brain markers at both past and future time points from any number of visits at any time. To overcome the data sparsity challenge, we construct spatial and temporal graphs and design the Temporal Spatial Layer (TSL) that adaptively aggregates information from all available observations. Comprehensive experiments on 287 typically developing term-born babies, comprising 680 longitudinal MRI scans collected between 2 weeks and 38 months of age in the Baby Connectome Project, demonstrate the robust performance of STRIDER. It reduces the mean absolute error (MAE) by 18.4%-54.6% compared with baseline methods while better preserving the high fidelity to the real data. The imputed trajectories further yield substantial downstream benefits, increasing the accuracy of cognitive prediction by 46.5%-110.7%. Crucially, the brain age gap (BAG) derived from STRIDER-imputed data exhibits a stronger correlation and prediction power for autism spectrum disorder (ASD) risk, improving prediction accuracy by 48.7%-101.3%. Finally, STRIDER demonstrates good generalizability when evaluated on an unseen external cohort. These findings establish STRIDER as an effective solution for reconstructing complete and heterogeneous early brain developmental trajectories from sparse longitudinal data, offering significant potential for advancing neurodevelopmental research and downstream applications.