A Graph Transformer-Based method for Alzheimer’s Disease Prediction

کد مقاله : 1164-NAEC

نویسندگان

فخرالسادات میرشریفی * دانشگاه الزهرا

چکیده مقاله

Alzheimer’s Disease is a neurodegenerative condition characterized by progressive cognitive decline. Early and accurate prediction of AD is crucial for timely intervention, yet remains challenging due to complex brain connectivity alterations and subtle early-stage biomarkers. In this paper, we present a novel Graph Transformer framework for Alzheimer’s disease prediction that synergistically integrates graph convolutional networks and Transformer self-attention to capture both local anatomical connectivity and global functional interactions among brain regions. The model constructs subject-specific brain graphs from multimodal neuroimaging data and demographic information from the ADNI cohort, enabling simultaneous modeling of anatomical and functional interactions. Our Graph Transformer achieves a classification accuracy of 93.4% on AD versus cognitively normal or CN subjects, outperforming conventional 3D CNNs (92.9%) and prior CNN–GCN hybrid models (91.6%). Interpretation of learned attention weights highlights key regions of interest, notably the hippocampus and entorhinal cortex, and post-hoc clustering of attention patterns reveals putative AD subtypes. The model generalizes to clinically relevant and offers promise for extension to longitudinal and multi-site neuroimaging studies. Overall, our approach provides a high-performance, interpretable framework with strong translational potential as a clinical diagnostic tool for AD, advancing explainable AI in neuroimaging.

کلیدواژه ها

Keywords: Alzheimer's Prediction; Feature Extraction; Graph Neural Networks; Transformer Architecture

وضعیت: پذیرفته شده