Classification of Alzheimer’s Disease using Multilayer Graph Neural Networks and Multi-Modal Neuroimaging

Abstract

Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder primarily affecting older adults, leading to cognitive decline, memory loss, and impaired daily function. Given its rising prevalence and lack of a cure, early diagnosis is critical for timely intervention. Traditional diagnostic methods, relying on clinical evaluations or single modality neuroimaging, often fail to capture the complex brain alterations in AD. This study introduces a novel framework combining a Multilayer Graph Attention Network (GAT) with multimodal neuroimaging for robust AD classification. Structural and functional brain networks were constructed using diffusion MRI (dMRI) and functional MRI (fMRI), focusing on 80 cortical and subcortical brain regions as network nodes. A multilayer graph was formed by integrating intra and inter connectivities into a comprehensive supra-adjacency matrix, which effectively captures both anatomical and functional in teractions across brain regions. The supra-adjacency matrix is a sophisticated representation that combines connectivity data from multiple modalities into a single, unified structure, enabling the model to analyze complex relationships between brain regions across different network layers. The GAT model, designed to analyze this multilayer network, assigns adaptive attention weights to nodes and edges, emphasizing regions and connections most impacted by AD pathology. Trained and evaluated on a balanced dataset of 100 AD patients and 100 Normal Cognition subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), The model achieved a classification accuracy of 97.5%, outperforming with traditional single-modality meth ods and other machine learning models, despite challenges in collecting both types of brain imaging modalities for the same subjects. This research highlights the benefits of integrating multimodal neuroimaging data with advanced graph-based deep learning architectures to enhance diagnostic precision. In the future, we plan to improve model performance by incorporating Explainable Artificial Intelligence (XAI).