Title: Transformer-Guided Adaptive Diffusion in Multi-Modal Graph Neural Networks for Preclinical Alzheimer’s Disease Classification

Abstract:

Representing the brain as a graph provides essential insights for the diagnosis and prognosis of neurodegenerative disorders by analyzing connections between regions of interest (ROIs). Although Graph Neural Networks (GNNs) have recently emerged to effectively model this relational data, they face inherent constraints in interpreting brain networks. Convolutional techniques often fail to adequately aggregate information from distant neighborhoods, whereas attention-based mechanisms struggle to capture node-centric details, specifically in preserving vital features from key nodes. These limitations hinder the ability to detect disease-specific variations across diverse multi-modal features. To address this, we introduce an integrated framework that directs the diffusion process at each node through a downstream transformer. This approach aggregates both short- and long-range graph properties using diffusion kernels and multi-head attention, respectively. Our model demonstrates superior performance in classifying pre-clinical Alzheimer’s disease (AD) across various modalities. Furthermore, it successfully pinpoints critical ROIs linked to the preclinical stages of AD, highlighting its significant potential for early diagnosis and disease prediction.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC