An explainable adaptive graph learning framework for dynamic brain connectivity analysis in autism diagnosis

Autism Spectrum Disorder (ASD) is increasingly understood through the lens of large-scale brain connectivity, yet most computational approaches fail to capture the dynamic, subject-specific neural interactions that define it. Recent graph neural network (GNN) methods have incorporated adaptive graph formulations. The existing approaches typically update edge weights via a single attention-reweighting pass per layer, without recursive feedback between updated edges and node representations, and rely on linear amplitude-based connectivity measures that may miss the non-linear phase synchrony abnormalities characteristic of ASD. We present the Adaptive Edge–Node Interaction Network (AENIN), a graph learning framework in which node representations and functional edges co-evolve through an iterative feedback mechanism, enabling connectivity to emerge from the data. Subject-specific brain graphs are constructed via Adaptive Phase-Synchronisation (APS), a method designed to extract temporally sensitive inter-regional dynamics from resting-state fMRI. In addition, a clinically interpretable Supportive Clinical Reporting Module (SCRM) proposed here bridges learned graph representations and clinical insight by quantifying functional connectivity patterns through three metrics: the Critical Influence Score (CIS), the Interaction Topology Score (ITS), and the Path Connectivity Score (PCS). The model evaluated on 800 participants from the ABIDE I dataset, AENIN achieves 0.8275 ± 0.0457 accuracy and 0.8304 ± 0.0462 F1-score. The neurobiological analysis reveals atypical hyperconnectivity in sensorimotor and occipital regions in ASD, corroborating findings in the existing literature.