ShiPMAN: A Self-Learnable Sparse Hierarchical Pattern and Multi-Attention Network for Interpretable Brain Connectivity in Autism

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition associated with atypical functional brain organization. To capture its network-level alterations, we propose an interpretable and sparse graph-based framework for brain connectivity analysis using resting-state fMRI (rs-fMRI). We investigate the large-scale functional organization of the brain and relate graph-level alterations to functional integration and segregation relevant to ASD cognitive phenotypes. Our method begins with rigorous preprocessing and brain parcellation, followed by z-score normalization of time series. Temporal dependencies across brain regions are modeled using Bidirectional Gated Recurrent Units (Bi-GRUs), enabling rich embedding of region-wise dynamics. These embeddings are used to compute pairwise attention scores, which are sparsified via top-\((k)\) filtering and \((\alpha)\)-entmax transformation, yielding biologically plausible brain graphs. Sparsification retains the most informative and biologically relevant connections, suppressing noise and enhancing interpretability. Applied to the ABIDE dataset, our framework reveals reduced clustering, lower efficiency, increased MST diameters, and altered motif patterns in ASD networks compared to the healthy brain. This framework integrates deep temporal modeling with explainable graph construction and multi-scale network analysis, offering a robust foundation for characterizing atypical connectivity patterns in ASD. Our findings highlight the promise of sparse attention-based graph modeling as a principled and interpretable approach to advancing functional connectomics in neurodevelopmental disorders.