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

Autism Spectrum Disorder (ASD) is known to exhibit a more rapid expansion in brain structure during the early years of life compared to typically developing children (TD). This is of utmost importance in understanding atypical brain development and forecasting the onset of ASD. However, the precise age-related cortical changes (trajectories) that could pinpoint atypical brain development in ASD remain largely unknown. In this study, we characterize the distinct developmental patterns of cortical morphology in individuals with ASD, investigate the important neural biomarkers for ASD diagnostics, and propose a deep-learning workflow that combined graph convolutional networks with low-rank multi-model tensor fusion (LMFGCN) for ASD prediction. Our findings reveal that the constituents of gray matter volume (GV), cortical thickness (CT), and surface area (SA) exhibit separable developmental trajectories in ASD. Furthermore, we identify regional differences in CT and SA that underscore the separable brain developmental trajectories, both of which contribute to changes in GV. Our study also demonstrates that LMFGCN, an end-to-end deep-learning model with intermediate integrative approach, outperforms early and late integration methods and other state-of-the-art models in ASD classification. Overall, our results highlight the importance of distinguishing between cortical SA and CT for understanding ASD pathobiology, particularly during the early brain overgrowth period, and demonstrate the potential utility of LMFGCN in ASD classification.