Title : Decoding autism heterogeneity: Dynamic semi-supervised clustering identifies patterns and pathways
Abstract:
Heterogeneity of Autism Spectrum Disorder (ASD) is a challenge for precision diagnosis or prognosis, thus motivating neurosubtyping and dimensional efforts. Nevertheless, most dimensional studies remain limited to identifying ASD-related patterns, especially using static brain connectivity. Lead to the dynamic nature of brain region signals is ignored, making it difficult to predict the progression pathways of ASD and hindering early intervention treatment. The improved Hilbert-Huang Transform (HHT) and Dynamic Semi-Supervised Clustering via Generative Adversarial Networks (DSmile-GAN) were combined to map ASD heterogeneity. Specifically, we first used an improved HHT to construct the dynamic brain network (DBN), focusing more on transient abnormalities in brain region signals of ASD participants. Then, DSmile-GAN was used to calculate the Betweenness Centrality (BC) and Degree Centrality (DC) of DBN, thereby identifying ASD-related patterns and disease progression using the differences between normal controls and ASD participants. Identification yielded three reproducible ASD-related patterns and two progression pathways: One characterized by visuomotor dysfunction and another by language-attentional deficits. Both pathways are associated with abnormal BC and DC in key brain regions and may serve as potential neurobiomarkers. The integration of improved HHT with DSmile-GAN provides a reliable method for analyzing the heterogeneity of ASD through DBN. Significantly, the two progression pathways in ASD offer guidance for targeted early interventions to mitigate functional impairments, thereby bridging the gap between neurobiological complexity and clinical practice.
