Distinct Neurodevelopmental Patterns and Intermediate Integration-Based Predictive Modeling in Autism

Abstract Autism Spectrum Disorder (ASD) is known to exhibit a more rapid brain expansion during early postnatal period of life, followed by a deceleration in age-related growth compared to typically developing children (TD). However, the development and evolution of the cerebral cortex that drive these changes remains unclear. In this study, we characterize the distinct developmental trajectory of cortical features in individuals with ASD across the lifespan, investigate the regional differences between ASD and TD, and propose a deep-learning workflow that combines graph convolutional networks with low-rank multi-model tensor fusion (LMFGCN) for ASD identification. Our findings reveal that the brain neurodevelopmental process in ASD is not a uniform process but rather is characterized by distinct developmental trajectories that vary across age and regions. The separated trajectories between cortical thickness (CT) and surface area (SA) especially at the early stage (about ages 0-8 years) indicate that early brain volume overgrowth in ASD is predominantly attributed to increased SA, rather than CT. The regional differences in CT and SA show spatially disrupted patterns that are congruent with changes in cortical volume. Moreover, we introduce LMFGCN, a deep-learning framework with intermediate fusion approach, outperforms early and late fusion methods and other state-of-the-art models in ASD classification. Overall, our results provide insights into spatiotemporal gradients of brain maturation, identify distinct brain alterations across the lifespan, and demonstrate the potential utility of LMFGCN in ASD identification.