Adaptive neighborhood aggregation graph neural network for early diagnosis of Alzheimer’s disease

Abstract

Early diagnosis of Alzheimer’s disease (AD) is crucial for timely intervention but remains challenging due to subtle and heterogeneous brain alterations, particularly in the mild cognitive impairment (MCI) stage. To address this issue, we propose a pathology-aware Adaptive Neighborhood Aggregation Graph Neural Network (ANA-GNN) to model the brain as a dynamic and task-driven graph for multimodal AD classification. The framework integrates three synergistic components: an adaptive neighborhood aggregation module that aligns each node’s receptive field with disease-specific heterogeneity, an importance-weighted pooling mechanism that enhances discriminative graph-level representations by prioritizing biologically relevant regions, and a gated multimodal fusion strategy that adaptively balances imaging and non-imaging information. Evaluated on an expanded cohort of 707 subjects from the ADNI dataset, ANA-GNN achieved an overall accuracy of 85.23% and an F1-score of 85.44%, consistently outperforming state-of-the-art baselines such as BrainGNN and Graph Transformers. Furthermore, the identified high-importance brain regions, including the hippocampus, amygdala, and posterior cingulate cortex, align with known AD biomarkers, demonstrating the model’s biological interpretability and potential as a reliable tool for early AD diagnosis.